We introduce a novel 3D generative method, Generative 3D Reconstruction (G3DR) in ImageNet, capable of generating diverse and high-quality 3D objects from single images,addressing the limitations of existing methods. At the heart of our framework is a novel depth regularization technique that enables the generation of scenes with high-geometric fidelity. G3DR also leverages a pre-trained language-vision model, such as CLIP, to enable reconstruction in novel views and improve the visual realism of generations. Additionally, G3DR designs a simple but effective sampling procedure to further improve the quality of generations.G3DR offers diverse and efficient 3D asset generation based on class or text conditioning. Despite its simplicity, G3DR is able to beat state-of-the-art methods, improving over them by up to 22% in perceptual metrics and 90% in geometry scores, while needing only half of the training time.We provide development code in the supplementary.

3D city generation is a desirable yet challenging task, since humans are more sensitive to structural distortions in urban environments. Additionally, generating 3D cities is more complex than 3D natural scenes since buildings, as objects of the same class, exhibit a wider range of appearances compared to the relatively consistent appearance of objects like trees in natural scenes. To address these challenges, we propose \textbf{CityDreamer}, a compositional generative model designed specifically for unbounded 3D cities. Our key insight is that 3D city generation should be a composition of different types of neural fields: 1) various building instances, and 2) background stuff, such as roads and green lands. Specifically, we adopt the bird's eye view scene representation and employ a volumetric render for both instance-oriented and stuff-oriented neural fields. The generative hash grid and periodic positional embedding are tailored as scene parameterization to suit the distinct characteristics of building instances and background stuff. Furthermore, we contribute a suite of CityGen Datasets, including OSM and GoogleEarth, which comprises a vast amount of real-world city imagery to enhance the realism of the generated 3D cities both in their layouts and appearances. CityDreamer achieves state-of-the-art performance not only in generating realistic 3D cities but also in localized editing within the generated cities.

Estimating the 6D object pose from a single RGB image often involves noise and indeterminacy due to challenges such as occlusions and cluttered backgrounds. Meanwhile, diffusion models have shown appealing performance in generating high-quality images from random noise with high indeterminacy through step-by-step denoising. Inspired by their denoising capability, we propose a novel diffusion-based framework (6D-Diff) to handle the noise and indeterminacy in object pose estimation for better performance. In our framework, to establish accurate 2D-3D correspondence, we formulate 2D keypoints detection as a reverse diffusion (denoising) process.To facilitate such a denoising process, we design a Mixture-of-Cauchy-based forward diffusion process and condition the reverse process on the object appearance features.Extensive experiments on the LM-O and YCB-V datasets demonstrate the effectiveness of our framework.

Social interaction is a fundamental aspect of human behavior and communication. The way individuals position themselves in relation to others, also known as proxemics, conveys social cues and affects the dynamics of social interaction. Reconstructing such interaction from images presents challenges because of mutual occlusion and the limited availability of large training datasets. To address this, we present a novel approach that learns a prior over the 3D proxemics of two people in close social interaction and demonstrate its use for single-view 3D reconstruction. We start by creating 3D training data of interacting people using image datasets with contact annotations. We then model the proxemics using a novel denoising diffusion model called BUDDI that learns the joint distribution over the poses of two people in close social interaction. Sampling from our generative proxemics model produces realistic 3D human interactions, which we validate through a perceptual study. We use BUDDI in reconstructing two people in close proximity from a single image without any contact annotation via an optimization approach that uses the diffusion model as a prior. Our approach recovers accurate and plausible 3D social interactions from noisy initial estimates, outperforming state-of-the-art methods.

Monocular depth estimation is a fundamental computer vision task. Recovering 3D depth from a single image is geometrically ill-posed and requires scene understanding, so it is not surprising that the rise of deep learning has led to a breakthrough. The impressive progress of monocular depth estimators has mirrored the growth in model capacity, from relatively modest CNNs to large Transformer architectures. Still, monocular depth estimators tend to struggle when presented with images with unfamiliar content and layout, since their knowledge of the visual world is restricted by the data seen during training, and challenged by zero-shot generalization to new domains. This motivates us to explore whether the extensive priors captured in recent generative diffusion models can enable better, more generalizable depth estimation. We introduce Marigold, a method for affine-invariant monocular depth estimation that is derived from Stable Diffusion and retains its rich prior knowledge. The estimator can be fine-tuned in a couple of days on a single GPU using only synthetic training data. It delivers state-of-the-art performance across a wide range of datasets, including over 20% performance gains in specific cases. Source code will be made publicly available.

We present MVD-Fusion: a method for single-view 3D inference via generative modeling of multi-view-consistent RGB-D images. While recent methods pursuing 3D inference advocate learning novel-view generative models, these generations are not 3D-consistent and require a distillation process to generate a 3D output. We instead cast the task of 3D inference as that of directly generating mutually-consistent multiple views, and we build on the insight that additionally inferring depth can provide a mechanism for enforcing this consistency. Specifically, we train a denoising diffusion model to generate multi-view RGB-D images given a single RGB input image and leverage the (intermediate noisy) depth estimates to obtain reprojection-based conditioning for enabling multi-view consistency. We train our system using a large-scale synthetic dataset, and evaluate our approach on both, synthetic and real-world data. We demonstrate that our approach can yield more accurate synthesis compared to recent state-of-the-art, including distillation-based 3D inference and prior multi-view generation methods. We also evaluate the geometry induced by our multi-view depth prediction and find that it yields a more accurate representation than other direct 3D inference approaches.

Three-dimensional (3D) reconstruction from a single image is an ill-posed problem with inherent ambiguities, i.e. scale. Predicting a 3D scene from text description(s) is similarly ill-posed, i.e. spatial arrangements of objects described. We investigate the question of whether two inherently ambiguous modalities can be used in conjunction to produce metric-scaled reconstructions. To test this, we focus on monocular depth estimation, the problem of predicting a dense depth map from a single image, but with an additional text caption describing the scene. To this end, we begin by encoding the text caption as a mean and standard deviation; using a variational framework, we learn the distribution of the plausible metric reconstructions of 3D scenes corresponding to the text captions as a prior. To ``select'' a specific reconstruction or depth map, we encode the given image through a conditional sampler that samples from the latent space of the variational text encoder, which is then decoded to the output depth map. Our approach is trained alternatingly between the text and image branches: in one optimization step, we predict the mean and standard deviation from the text description and sample from a standard Gaussian, and in the other, we sample using a (image) conditional sampler. Once trained, we directly predict depth from the encoded text using the conditional sampler. We demonstrate our approach on indoor (NYUv2) and outdoor (KITTI) scenarios, where we show that language can consistently improve performance in both. Code: https://github.com/Adonis-galaxy/WorDepth.

We introduce Free3D, a simple accurate method for monocular open-set novel view synthesis (NVS). Similar to Zero-1-to-3, we start from a pre-trained 2D image generator for generalization, and fine-tune it for NVS. Compared to other works that took a similar approach, we obtain significant improvements without resorting to an explicit 3D representation, which is slow and memory-consuming, andwithout training an additional network for 3D reconstruction. Our key contribution is to improve the way the target camera pose is encoded in the network, which we do by introducing a new ray conditioning normalization (RCN) layer. The latter injects pose information in the underlying 2D image generator by telling each pixel its viewing direction. We further improve multi-view consistency by using light-weight multi-view attention layers and by sharing generation noise between the different views. We train Free3D on the Objaverse dataset and demonstrate excellent generalization to new categories in new datasets, including OmniObject3D and GSO. The project page is available at https://chuanxiaz.com/free3d/.

Human Mesh Recovery (HMR) aims to estimate the 3D human body from 2D images, which is a challenging task due to inherent ambiguities in translating 2D observations to 3D space. A novel approach called PostureHMR is proposed to leverage a multi-step diffusion-style process, which converts this task into a posture transformation from an SMPL T-pose mesh to the target mesh. To inject the learning process of posture transformation with the physical structure of the human body model, a kinematics-based forward process is proposed to interpolate the intermediate state with pose and shape decomposition. Moreover, a mesh-to-posture (M2P) decoder is designed, by combining the input of 3D and 2D mesh constraints estimated from the image to model the posture changes in the reverse process. It mitigates the difficulties of posture change learning directly from RGB pixels. To overcome the limitation of pixel-level misalignment of modeling results with the input image, a new trimap-based rendering loss is designed to highlight the areas with poor recognition. Experiments conducted on three widely used datasets demonstrate that the proposed approach outperforms the state-of-the-art methods.

This paper introduces 3DFIRES, a novel system for scene-level 3D reconstruction from posed images. Designed to work with as few as one view, 3DFIRES reconstructs the complete geometry of unseen scenes, including hidden surfaces. With multiple view inputs, our method produces full reconstruction within all camera frustums. A key feature of our approach is the fusion of multi-view information at the feature level, enabling the production of coherent and comprehensive 3D reconstruction. We train our system on non-watertight scans from large-scale real scene dataset. We show it matches the efficacy of single-view reconstruction methods with only one input and surpasses existing techniques in both quantitative and qualitative measures for sparse-view 3D reconstruction.

Learning 3D models of all animals in nature requires massively scaling up existing solutions. With this ultimate goal in mind, we develop 3D-Fauna, an approach that learns a pan-category deformable 3D animal model for more than 100 animal species jointly. One crucial bottleneck of modeling animals is the limited availability of training data, which we overcome by learning our model from 2D Internet images. We show that prior approaches, which are category-specific, fail to generalize to rare species with limited training images. We address this challenge by introducing the Semantic Bank of Skinned Models (SBSM), which automatically discovers a small set of base animal shapes by combining geometric inductive priors with semantic knowledge implicitly captured by an off-the-shelf self-supervised feature extractor.To train such a model, we also contribute a new large-scale dataset of diverse animal species. At inference time, given a single image of any quadruped animal, our model reconstructs an articulated 3D mesh in a feed-forward manner in seconds.

Depth completion aims to derive a dense depth map from sparse depth measurements with a synchronized color image. Current state-of-the-art (SOTA) methods are predominantly propagation-based, which work as an iterative refinement on the initial estimated dense depth. However, the initial depth estimations mostly result from direct applications of convolutional layers on the sparse depth map. In this paper, we present a Bilateral Propagation Network (BP-Net), that propagates depth at the earliest stage to avoid directly convolving on sparse data. Specifically, our approach propagates the target depth from nearby depth measurements via a non-linear model, whose coefficients are generated through a multi-layer perceptron conditioned on both \emph{radiometric difference} and \emph{spatial distance}. By integrating bilateral propagation with multi-modal fusion and depth refinement in a multi-scale framework, our BP-Net demonstrates outstanding performance on both indoor and outdoor scenes. It achieves SOTA on the NYUv2 dataset and ranks 1st on the KITTI depth completion benchmark at the time of submission. Experimental results not only show the effectiveness of bilateral propagation but also emphasize the significance of early-stage propagation in contrast to the refinement stage. Our code and trained models will be available on the project page.

The recent success in revealing scene details from sparse 3D point clouds obtained via structure-from-motion has raised significant privacy concerns in visual localization. One prominent approach for mitigating this issue is to lift 3D points to 3D lines thereby reducing the effectiveness of the scene inversion attacks, but this comes at the cost of increased algorithmic complexity for camera localization due to weaker geometric constraints induced by line clouds. To overcome this limitation, we propose a new lifting approach called ``ray cloud'', whereby each lifted 3D line intersects at one of two predefined locations, depicting omnidirectional rays from two cameras. This yields two benefits, i) camera localization can now be cast as relative pose estimation between the query image and the calibrated rig of two perspective cameras which can be efficiently solved using a variant of the 5-point algorithm, and ii) the ray cloud introduces erroneous estimations for the density-based inversion attack, degrading the quality of scene recovery. Moreover, we explore possible modifications of the inversion attack to better recover scenes from the ray clouds and propose a ray sampling technique to reduce the effectiveness of the modified attack. Experimental results on two public datasets show real-time localization speed as well as enhanced privacy-preserving capability over the state-of-the-art without overly sacrificing the localization accuracy.

Generating multiview images from a single view facilitates the rapid generation of a 3D mesh conditioned on a single image. Recent methods that introduce 3D global representation into diffusion models have shown the potential to generate consistent multiviews, but they have reduced generation speed and face challenges in maintaining generalizability and quality. To address this issue, we propose EpiDiff, a localized interactive multiview diffusion model. At the core of the proposed approach is to insert a lightweight epipolar attention block into the frozen diffusion model, leveraging epipolar constraints to enable cross-view interaction among feature maps of neighboring views. The newly initialized 3D modeling module preserves the original feature distribution of the diffusion model, exhibiting compatibility with a variety of base diffusion models. Experiments show that EpiDiff generates 16 multiview images in just 12 seconds, and it surpasses previous methods in quality evaluation metrics, including PSNR, SSIM and LPIPS. Additionally, EpiDiff can generate a more diverse distribution of views, improving the reconstruction quality from generated multiviews.

In this paper, we democratise 3D content creation, enabling precise generation of 3D shapes from abstract sketches while overcoming limitations tied to drawing skills. We introduce a novel part-level modelling and alignment framework that facilitates abstraction modelling and cross-modal correspondence. Leveraging the same part-level decoder, our approach seamlessly extends to sketch modelling by establishing correspondence between CLIPasso edgemaps and projected 3D part regions, eliminating the need for a dataset pairing human sketches and 3D shapes. Additionally, our method introduces a seamless in-position editing process as a byproduct of cross-modal part-aligned modelling. Operating in a low-dimensional implicit space, our approach significantly reduces computational demands and processing time.

In this paper, we present a tensor decomposition and low-rank recovery approach (LowRankOcc) for vision-based 3D semantic occupancy prediction. Conventional methods model outdoor scenes with fine-grained 3D grids, but the sparsity of non-empty voxels introduces considerable spatial redundancy, leading to potential overfitting risks. In contrast, our approach leverages the intrinsic low-rank property of 3D occupancy data, factorizing voxel representations into low-rank components to efficiently mitigate spatial redundancy without sacrificing performance. Specifically, we present the Vertical-Horizontal (VH) decomposition block factorizes 3D tensors into vertical vectors and horizontal matrices. With our "decomposition-encoding-recovery" framework, we encode 3D contexts with only 1/2D convolutions and poolings, and subsequently recover the encoded compact yet informative context features back to voxel representations. Experimental results demonstrate that LowRankOcc achieves state-of-the-art performances in semantic scene completion on the SemanticKITTI dataset and 3D occupancy prediction on the nuScenes dataset.

CNC manufacturing is a process that employs computer numerical control (CNC) machines to govern the movements of various industrial tools and machinery, encompassing equipment ranging from grinders and lathes to mills and CNC routers. However, the reliance on manual CNC programming has become a bottleneck, and the requirement for expert knowledge can result in significant costs. Therefore, we introduce a pioneering approach named CNC-Net, representing the use of deep neural networks (DNNs) to simulate CNC machines and grasp intricate operations when supplied with raw materials. CNC-Net constitutes a self-supervised framework that exclusively takes an input 3D model and subsequently generates the essential operation parameters required by the CNC machine to construct the object. Our method has the potential to transformative automation in manufacturing by offering a cost-effective alternative to the high costs of manual CNC programming while maintaining exceptional precision in 3D object production. Our experiments underscore the effectiveness of our CNC-Net in constructing the desired 3D objects through the utilization of CNC operations. Notably, it excels in preserving finer local details, exhibiting a marked enhancement in precision compared to the state-of-the-art 3D CAD reconstruction approaches. The codes are available athttps://github.com/myavartanoo/CNC-Net_PyTorch.

We present an approach that can reconstruct hands in 3D from monocular input. Our approach for Hand Mesh Recovery, HaMeR, follows a fully transformer-based architecture and can analyze hands with significantly increased accuracy and robustness compared to previous work. The key to HaMeR’s success lies in scaling up both the data used for training and the capacity of the deep network for hand reconstruction. For training data, we combine multiple datasets that contain 2D or 3D hand annotations. For the deep model, we use a large scale Vision Transformer architecture. Our final model consistently outperforms the previous baselines on popular 3D hand pose benchmarks. To further evaluate the effect of our design in non-controlled settings, we annotate existing in-the-wild datasets with 2D hand keypoint annotations. On this newly collected dataset of annotations, HInt, we demonstrate significant improvements over existing baselines. We will make our code, data and models publicly available upon publication.

Inferring scene geometry from images via Structure from Motion is a long-standing and fundamental problem in computer vision. While classical approaches and, more recently, depth map predictions only focus on the visible parts of a scene, the task of scene completion aims to reason about geometry even in occluded regions. With the popularity of neural radiance fields (NeRFs), implicit representations also became popular for scene completion by predicting so-called density fields. Unlike explicit approaches e.g. voxel-based methods, density fields also allow for accurate depth prediction and novel-view synthesis via image-based rendering. In this work, we propose to fuse the scene reconstruction from multiple images and distill this knowledge into a more accurate single-view scene reconstruction. To this end, we propose Multi-View Behind the Scenes (MVBTS) to fuse density fields from multiple posed images, trained fully self-supervised only from image data. Using knowledge distillation, we use MVBTS to train a single-view scene completion network via direct supervision called KDBTS. It achieves state-of-the-art performance on occupancy prediction, especially in occluded regions.

Recovering the 3D scene geometry from a single view is a fundamental yet ill-posed problem in computer vision. While classical depth estimation methods infer only a 2.5D scene representation limited to the image plane, recent approaches based on radiance fields reconstruct a full 3D representation. However, these methods still struggle with occluded regions since inferring geometry without visual observation requires (i) semantic knowledge of the surroundings, and (ii) reasoning about spatial context. We propose KYN, a novel method for single-view scene reconstruction that reasons about semantic and spatial context to predict each point's density. We introduce a vision-language modulation module to enrich point features with fine-grained semantic information. We aggregate point representations across the scene through a language-guided spatial attention mechanism to yield per-point density predictions aware of the 3D semantic context. We show that KYN improves 3D shape recovery compared to predicting density for each 3D point in isolation. We achieve state-of-the-art results in scene and object reconstruction on KITTI-360, and show improved zero-shot generalization compared to prior work. Project page: https://ruili3.github.io/kyn.

Dense depth maps have been used as a key element of visual perception tasks. There have been tremendous efforts to enhance the depth quality, ranging from optimization-based to learning-based methods. Despite the remarkable progress for a long time, their applicability in the real world is limited due to systematic measurement biases such as density, sensing pattern, and scan range. It is well-known that the biases make it difficult for these methods to achieve their generalization. We observe that learning a joint representation for input modalities (e.g., images and depth), which most recent methods adopt, is sensitive to the biases. In this work, we disentangle those modalities to mitigate the biases with prompt engineering. For this, we design a novel depth prompt module to allow the desirable feature representation according to new depth distributions from either sensor types or scene configurations. Our depth prompt can be embedded into foundation models for monocular depth estimation. Through this embedding process, our method helps the pretrained model to be free from restraint of depth scan range and to provide absolute scale depth maps. We demonstrate the effectiveness of our method through extensive evaluations. Source code is publicly available at https://github.com/JinhwiPark/DepthPrompting.

Novel-view synthesis through diffusion models has demonstrated remarkable potential for generating diverse and high-quality images. Yet, the independent process of image generation in these prevailing methods leads to challenges in maintaining multiple view consistency. To address this, we introduce ViewFusion, a novel, training-free algorithm that can be seamlessly integrated into existing pre-trained diffusion models. Our approach adopts an auto-regressive method that implicitly leverages previously generated views as context for next view generation, ensuring robust multi-view consistency during the novel-view generation process. Through a diffusion process that fuses known-view information via interpolated denoising, our framework successfully extends single-view conditioned models to work in multiple-view conditional settings without any additional fine-tuning. Extensive experimental results demonstrate the effectiveness of ViewFusion in generating consistent and detailed novel views.

We introduce multi-slice reasoning, a new notion for single-view 3D reconstruction which challenges the current and prevailing belief that multi-view synthesis is the most natural conduit between single-view and 3D. Our key observation is that object slicing is more advantageous than altering views to reveal occluded structures. Specifically, slicing is more occlusion-revealing since it can peel through any occluders without obstruction. In the limit, i.e., with infinitely many slices, it is guaranteed to unveil all hidden object parts. We realize our idea by developing Slice3D, a novel method for single-view 3D reconstruction which first predicts multi-slice images from a single RGB image and then integrates the slices into a 3D model using a coordinate-based transformer network for signed distance prediction. The slice images can be regressed or generated, both through a U-Net based network. For the former, we inject a learnable slice indicator code to designate each decoded image into a spatial slice location, while the slice generator is a denoising diffusion model operating on the entirety of slice images stacked on the input channels. We conduct extensive evaluation against state-of-the-art alternatives to demonstrate superiority of our method, especially in recovering complex and severely occluded shape structures, amid ambiguities. All Slicer3D results were produced by networks trained on a single Nvidia A40 GPU, with an inference time less than 20 seconds.

Score distillation sampling (SDS) and its variants have greatly boosted the development of text-to-3D generation, but are vulnerable to geometry collapse and poor textures yet. To solve this issue, we first deeply analyze the SDS and find that its distillation sampling process indeed corresponds to the trajectory sampling of a stochastic differential equation (SDE): SDS samples along an SDE trajectory to yield a less noisy sample which then serves as a guidance to optimize a 3D model. However, the randomness in SDE sampling often leads to a diverse and unpredictable sample which is not always less noisy, and thus is not a consistently correct guidance, explaining the vulnerability of SDS. Since for any SDE, there always exists an ordinary differential equation (ODE) whose trajectory sampling can deterministically and consistently converge to the desired target point as the SDE, we propose a novel and effective ``Consistent3D" method that explores the ODE deterministic sampling prior for text-to-3D generation. Specifically, at each training iteration, given a rendered image by a 3D model, we first estimate its desired 3D score function by a pre-trained 2D diffusion model, and build an ODE for trajectory sampling. Next, we design a consistency distillation sampling loss which samples along the ODE trajectory to generate two adjacent samples and uses the less noisy sample to guide another more noisy one for distilling the deterministic prior into the 3D model. Experimental results show the efficacy of our Consistent3D in generating high-fidelity and diverse 3D objects and large-scale scenes, as shown in Fig. 1. Code and models will be released online.

We present GigaPose, a fast, robust, and accurate method for CAD-based novel object pose estimation in RGB images. GigaPose first leverages discriminative templates, rendered images of the CAD models, to recover the out-of-plane rotation and then uses patch correspondences to estimate the four remaining parameters. Our approach samples templates in only a two-degrees-of-freedom space instead of the usual three and matches the input image to the templates using fast nearest-neighbor search in feature space, results in a speedup factor of 38x compared to the state of the art. Moreover, GigaPose is significantly more robust to segmentation errors. Our extensive evaluation on the seven core datasets of the BOP challenge demonstrates that it achieves state-of-the-art accuracy and can be seamlessly integrated with a refinement method. Additionally, we show the potential of GigaPose with 3D models predicted by recent work on 3D reconstruction from a single image, relaxing the need for CAD models and making 6D pose object estimation much more convenient. We will make the code and trained models publicly available.

Lifting 2D diffusion for 3D generation is a challenging problem due to the lack of geometric prior and the complex entanglement of materials and lighting in natural images. Existing methods have shown promise by first creating the geometry through score-distillation sampling (SDS) applied to rendered surface normals, followed by appearance modeling. However, relying on a 2D RGB diffusion model to optimize surface normals is suboptimal due to the distribution discrepancy between natural images and normals maps, leading to instability in optimization. In this paper, recognizing that the normal and depth information effectively describe scene geometry and be automatically estimated from images, we propose to learn a generalizable Normal-Depth diffusion model for 3D generation. We achieve this by training on the large-scale LAION dataset together with the generalizable image-to-depth and normal prior models. In an attempt to alleviate the mixed illumination effects in the generated materials, we introduce an albedo diffusion model to impose data-driven constraints on the albedo component. Our experiments show that when integrated into existing text-to-3D pipelines, our models significantly enhance the detail richness, achieving state-of-the-art results. Our project page is at https://aigc3d.github.io/richdreamer/.

360 depth estimation has recently received great attention for 3D reconstruction owing to its omnidirectional field of view (FoV).Recent approaches are predominantly focused on cross-projection fusion with geometry-based re-projection: they fuse 360 images with equirectangular projection (ERP) and another projection type, e.g., cubemap projection to estimate depth with the ERP format. However, these methods suffer from 1) limited local receptive fields, making it hardly possible to capture large FoV scenes, and 2) prohibitive computational cost, caused by the complex cross-projection fusion module design. In this paper, we propose $Elite360D$, a novel framework that inputs the ERP image and icosahedron projection (ICOSAP) point set, which is undistorted and spatially continuous. Elite360D is superior in its capacity in learning a representation from a local-with-global perspective. With a flexible ERP image encoder, it includes an ICOSAP point encoder, and a Bi-projection Bi-attention Fusion (B2F) module (totally ~1M parameters). Specifically, the ERP image encoder can take various perspective image-trained backbones (e.g., ResNet, Transformer) to extract local features. The point encoder extracts the global features from the ICOSAP. Then, the B2F module captures the semantic- and distance-aware dependencies between each pixel of the ERP feature and the entire ICOSAP feature set. Without specific backbone design and obvious computational cost increase, Elite360D outperforms the prior arts on several benchmark datasets.

Creating high-quality 3D models of clothed humans from single images for real-world applications is crucial. Despite recent advancements, accurately reconstructing humans in complex poses or with loose clothing from in-the-wild images, along with predicting textures for unseen areas, remains a significant challenge. A key limitation of previous methods is their insufficient prior guidance in transitioning from 2D to 3D and in texture prediction. In response, we introduce \textbf{SIFU} (\textbf{S}ide-view Conditioned \textbf{I}mplicit \textbf{F}unction for Real-world \textbf{U}sable Clothed Human Reconstruction), a novel approach combining a Side-view Decoupling Transformer with a 3D Consistent Texture Refinement pipeline. SIFU employs a cross-attention mechanism within the transformer, using SMPL-X normals as queries to effectively decouple side-view features in the process of mapping 2D features to 3D. This method not only improves the precision of the 3D models but also their robustness, especially when SMPL-X estimates are not perfect. Our texture refinement process leverages text-to-image diffusion-based prior to generate realistic and consistent textures for invisible views. Through extensive experiments, SIFU surpasses SOTA methods in both geometry and texture reconstruction, showcasing enhanced robustness in complex scenarios and achieving an unprecedented Chamfer and P2S measurement. Our approach extends to practical applications such as 3D printing and scene building, demonstrating its broad utility in real-world scenarios. Codes will be made available upon acceptance.

Score distillation sampling (SDS) has been widely adopted to overcome the absence of unseen views in reconstructing 3D objects from a single mage. It leverages pre-trained 2D diffusion models as teacher to guide the reconstruction of student 3D models. Despite its remarkable success, SDS-based methods often encounter geometric artifacts and texture saturation. We find out the crux is the overlooked indiscriminate treatment of diffusion time-steps during optimization: it unreasonably treats the student-teacher knowledge distillation to be equal at all time-steps and thus entangles coarse-grained and fine-grained modeling. Therefore, we propose the Diffusion Time-step Curriculum one-image-to-3D pipeline (DTC123), which involves both the teacher and student models collaborating with the time-step curriculum in a coarse-to-fine manner. Extensive experiments on NeRF4, RealFusion15, and Level50 benchmark demonstrate that DTC123 can produce multi-view consistent, high-quality, and diverse 3D assets.

Category-level object pose estimation, aiming to predict the 6D pose and 3D size of objects from known categories, typically struggles with large intra-class shape variation.Existing works utilizing mean shapes often fall short of capturing this variation.To address this issue, we present SecondPose, a novel approach integrating object-specific geometric features with semantic category priors from DINOv2.Leveraging the advantage of DINOv2 in providing SE(3)-consistent semantic features, we hierarchically extract two types of SE(3)-invariant geometric features to further encapsulate local-to-global object-specific information.These geometric features are then point-aligned with DINOv2 features to establish a consistent object representation under SE(3) transformations, facilitating the mapping from camera space to the pre-defined canonical space, thus further enhancing pose estimation.Extensive experiments on NOCS-REAL275 demonstrate that SecondPose achieves a 12.4\% leap forward over the state-of-the-art.Moreover, on a more complex dataset HouseCat6D which provides photometrically challenging objects, SecondPose still surpasses other competitors by a large margin.The code will be released soon.

In this work, we introduce Wonder3D, a novel method for generating high-fidelity textured meshes from single-view images with remarkable efficiency. Recent methods based on the Score Distillation Sampling (SDS) loss methods have shown the potential to recover 3D geometry from 2D diffusion priors, but they typically suffer from time-consuming per-shape optimization and inconsistent geometry. In contrast, certain works directly produce 3D information via fast network inferences, but their results are often of low quality and lack geometric details. To holistically improve the quality, consistency, and efficiency of image-to-3D tasks, we propose a cross-domain diffusion model that generates multi-view normal maps and the corresponding color images. To ensure consistency, we employ a multi-view cross-domain attention mechanism that facilitates information exchange across views and modalities. Lastly, we introduce a geometry-aware normal fusion algorithm that extracts high-quality surfaces from the multi-view 2D representations in only 2~3 minutes. Our extensive evaluations demonstrate that our method achieves high-quality reconstruction results, robust generalization, and remarkable efficiency compared to prior works.

We present En3D, an enhanced generative scheme for sculpting high-quality 3D human avatars. Unlike previous works that rely on scarce 3D datasets or limited 2D collections with imbalanced viewing angles and imprecise pose priors, our approach aims to develop a zero-shot 3D generative scheme capable of producing visually realistic, geometrically accurate and content-wise diverse 3D humans without relying on pre-existing 3D or 2D assets. To address this challenge, we introduce a meticulously crafted workflow that implements accurate physical modeling to learn the enhanced 3D generative model from synthetic 2D data. During inference, we integrate optimization modules to bridge the gap between realistic appearances and coarse 3D shapes. Specifically, En3D comprises three modules: a 3D generator that accurately models generalizable 3D humans with realistic appearance from synthesized balanced, diverse, and structured human images; a geometry sculptor that enhances shape quality using multi-view normal constraints for intricate human anatomy; and a texturing module that disentangles explicit texture maps with fidelity and editability, leveraging semantical UV partitioning and a differentiable rasterizer. Experimental results show that our approach significantly outperforms prior works in terms of image quality, geometry accuracy, and content diversity. We also showcase the applicability of our generated avatars for animation and editing, as well as the scalability of our approach for content-style free adaptation.

Previous works concerning single-view hand-held object reconstruction typically rely on supervision from 3D ground-truth models, which are hard to collect in real world. In contrast, readily accessible hand-object videos offer a promising training data source, but they only give heavily occluded object observations. In this paper, we present a novel synthetic-to-real framework to exploit Multi-view Occlusion-aware supervision from hand-object videos for Hand-held Object reconstruction (MOHO) from a single image, tackling two predominant challenges in such setting: hand-induced occlusion and object's self-occlusion. First, in the synthetic pre-training stage, we render a large-scaled synthetic dataset SOMVideo with hand-object images and multi-view occlusion-free supervisions, adopted to address hand-induced occlusion in both 2D and 3D spaces. Second, in the real-world finetuning stage, MOHO leverages the amodal-mask-weighted geometric supervision to mitigate the unfaithful guidance caused by the hand-occluded supervising views in real world. Moreover, domain-consistent occlusion-aware features are amalgamated in MOHO to resist object's self-occlusion for inferring the complete object shape. Extensive experiments on HO3D and DexYCB datasets demonstrate 2D-supervised MOHO gains superior results against 3D-supervised methods by a large margin.

Reconstructing human-object interaction in 3D from a single RGB image is a challenging task and existing data driven methods do not generalize beyond the objects present in the carefully curated 3D interaction datasets.Capturing large-scale real data to learn strong interaction and 3D shape priors is very expensive due to the combinatorial nature of human-object interactions. In this paper, we propose ProciGen (Procedural interaction Generation), a method to procedurally generate datasets with both, plausible interaction and diverse object variation.We generate 1M+ human-object interaction pairs in 3D and leverage this large-scale data to train our HDM (Hierarchical Diffusion Model), a novel method to reconstruct interacting human and unseen objects, without any templates. Our HDM is an image-conditioned diffusion model that learns both realistic interaction and highly accurate human and object shapes.Experiments show that our HDM trained with ProciGen significantly outperforms prior methods that requires template meshes and that our dataset allows training methods with strong generalization ability to unseen object instances. Our code and data will be publicly released.

Single image depth estimation is a foundational task in computer vision and generative modeling. However, prevailing depth estimation models grapple with accommodating the increasing resolutions commonplace in today's consumer cameras and devices. Existing high-resolution strategies show promise, but they often face limitations, ranging from error propagation to the loss of high-frequency details. We present $\textbf{PatchFusion}$, a novel tile-based framework with three key components to improve the current state of the art: $\textbf{(1)}$ A patch-wise fusion network that fuses a globally-consistent coarse prediction with finer, inconsistent tiled predictions via high-level feature guidance, $\textbf{(2)}$ A Global-to-Local (G2L) module that adds vital context to the fusion network, discarding the need for patch selection heuristics, and $\textbf{(3)}$ A Consistency-Aware Training (CAT) and Inference (CAI) approach, emphasizing patch overlap consistency and thereby eradicating the necessity for post-processing. Experiments on UnrealStereo4K, MVS-Synth, and Middleburry 2014 demonstrate that our framework can generate high-resolution depth maps with intricate details. PatchFusion is independent of the base model for depth estimation. Notably, our framework built on top of SOTA ZoeDepth brings improvements for a total of 17.3% and 29.4% in terms of the root mean squared error (RMSE) on UnrealStereo4K and MVS-Synth, respectively.

We present SPAD, a novel approach for creating consistent multi-view images from text prompts or single images. To enable multi-view generation, we repurpose a pretrained 2D diffusion model by extending its self-attention layers with cross-view interactions, and fine-tune it on a high quality subset of Objaverse. We find that a naive extension of the self-attention proposed in prior work (e.g., MVDream) leads to content copying between views. Therefore, we explicitly constrain the cross-view attention based on epipolar geometry. To further enhance 3D consistency, we utilize Pl ̈ucker coordinates derived from camera rays and inject them as positional encoding. This enables SPAD to reason over spatial proximity in 3D well. Compared to concurrent works that can only generate views at fixed azimuth and elevation (e.g., MVDream, SyncDreamer), SPAD offers full camera control and achieves state-of-the-art results in novel view synthesis on unseen objects from the Objaverse and Google Scanned Objects datasets. Finally, we demonstrate that text-to-3D generation using SPAD prevents the multi-face Janus issue.

Despite the progress of learning-based methods for 6D object pose estimation, the dilemma of the trade-off between accuracy and scalability for novel objects still exists.Specifically, previous methods for novel objects could not make good use of the target object's 3D shape information since they focus on generalization by processing the shape indirectly, making them less effective.We present GenFlow, an approach that enables both accuracy and generalization to novel objects by the guidance of the target object's shape.Our method predicts optical flow between the rendered image and the observed image and refines the 6D pose iteratively.It boosts the performance by a constraint of the 3D shape and the generalizable geometric knowledge learned from an end-to-end differentiable system.We further improve our model by designing a cascade network architecture to exploit the multi-scale correlations and coarse-to-fine refinement.GenFlow ranked first on the unseen object pose estimation benchmarks in both the RGB and RGB-D cases.It also achieves performance competitive with existing state-of-the-art methods for the seen object pose estimation without any fine-tuning.

We present PointInfinity, an efficient family of point cloud diffusion models. Our core idea is to use a transformer-based architecture with a fixed-size, resolution-invariant latent representation. This enables efficient training with low-resolution point clouds, while allowing high-resolution point clouds to be generated during inference. More importantly, we show that scaling the test-time resolution beyond the training resolution improves the fidelity of generated point clouds and surfaces. We analyze this phenomenon and draw a link to classifier-free guidance commonly used in diffusion models, demonstrating that both allow trading off fidelity and variability during inference. Experiments on CO3D show that PointInfinity can efficiently generate high-resolution point clouds (up to 131k points, 31 times more than Point-E) with state-of-the-art quality.

We study the problem of single-image zero-shot 3D shape reconstruction. Recent works learn zero-shot shape reconstruction through generative modeling of 3D assets, but these models are computationally expensive at train and inference time. In contrast, the traditional approach to this problem is regression-based, where deterministic models are trained to directly regress the object shape. Such regression methods possess much higher computational efficiency than generative methods. This raises a natural question: is generative modeling necessary for high performance, or conversely, are regression-based approaches still competitive? To answer this, we design a strong regression-based model, called ZeroShape, based on the converging findings in this field and a novel insight. We also curate a large real-world evaluation benchmark, with objects from three different real-world 3D datasets. This evaluation benchmark is more diverse and an order of magnitude larger than what prior works use to quantitatively evaluate their models, aiming at reducing the evaluation variance in our field. We show that ZeroShape not only achieves superior performance over state-of-the-art methods, but also demonstrates significantly higher computational and data efficiency.

Recent advancements in open-world 3D object generation have been remarkable, with image-to-3D methods offering superior fine-grained control over their text-to-3D counterparts. However, most existing models fall short in simultaneously providing rapid generation speeds and high fidelity to input images - two features essential for practical applications. In this paper, we present One-2-3-45++, an innovative method that transforms a single image into a detailed 3D textured mesh in approximately one minute. Our approach aims to fully harness the extensive knowledge embedded in 2D diffusion models and priors from valuable yet limited 3D data. This is achieved by initially finetuning a 2D diffusion model for consistent multi-view image generation, followed by elevating these images to 3D with the aid of multi-view-conditioned 3D native diffusion models. Extensive experimental evaluations demonstrate that our method can produce high-quality, diverse 3D assets that closely mirror the original input image.

In this work, we propose a method to address the challenge of rendering a 3D human from a single image in a free-view manner. Some existing approaches could achieve this by using generalizable pixel-aligned implicit fields to reconstruct a textured mesh of a human or by employing a 2D diffusion model as guidance with the Score Distillation Sampling (SDS) method, to lift the 2D image into 3D space. However, a generalizable implicit field often results in an over-smooth texture field, while the SDS method tends to lead to a texture-inconsistent novel view with the input image. In this paper, we introduce a texture-consistent back view synthesis method that could transfer the reference image content to the back view through depth-guided mutual self-attention. With this method, we could achieve high-fidelity and texture-consistent human rendering from a single image. Moreover, to alleviate the color distortion that occurs in the side region, we propose a \xy{visibility-aware patch consistency regularization} combined with the synthesized back view texture. Experiments conducted on both real and synthetic data demonstrate the effectiveness of our method and show that our approach outperforms previous baseline methods.

Recent learning methods for object pose estimation require resource-intensive training for each individual object instance or category, hampering their scalability in real applications when confronted with previously unseen objects. In this paper, we propose MatchU, a Fuse-Describe-Match strategy for 6D pose estimation from RGB-D images. MatchU is a generic approach that fuses 2D texture and 3D geometric cues for 6D pose prediction of unseen objects. We rely on learning geometric 3D descriptors that are rotation-invariant by design. By encoding pose-agnostic geometry, the learned descriptors naturally generalize to unseen objects and capture symmetries. To tackle ambiguous associations using 3D geometry only, we fuse additional RGB information into our descriptor. This is achieved through a novel attention-based mechanism that fuses cross-modal information, together with a matching loss that leverages the latent space learned from RGB data to guide the descriptor learning process. Extensive experiments reveal the generalizability of both the RGB-D fusion strategy as well as the descriptor efficacy. Benefiting from the novel designs, MatchU surpasses all existing methods by a significant margin in terms of both accuracy and speed, even without the requirement of expensive re-training or rendering.

Accurate monocular metric depth estimation (MMDE) is crucial to solving downstream tasks in 3D perception and modeling.However, the remarkable accuracy of recent MMDE methods is confined to their training domains.These methods fail to generalize to unseen domains even in the presence of moderate domain gaps, which hinders their practical applicability.We propose a new model, UniDepth, capable of reconstructing metric 3D scenes from solely single images across domains.Departing from the existing MMDE methods, UniDepth directly predicts metric 3D points from the input image at inference time without any additional information, striving for a universal and flexible MMDE solution.In particular, UniDepth implements a self-promptable camera module predicting dense camera representation to condition depth features.Our model exploits a pseudo-spherical output representation, which disentangles camera and depth representations.In addition, we propose a geometric invariance loss that promotes the invariance of camera-prompted depth features.Thorough evaluations on ten datasets in a zero-shot regime consistently demonstrate the superior performance of UniDepth, even when compared with methods directly trained on the testing domains. Code and models are available at: github.com/lpiccinelli-eth/unidepth.

Novel view synthesis aims to generate new view images of a given view image collection. Recent attempts address this problem relying on 3D geometry priors (e.g., shapes, sizes, and positions) learned from multi-view images. However, such methods encounter the following limitations: 1) they require a set of multi-view images as training data for a specific scene (e.g., car or chair), which is often unavailable in many real-world scenarios; 2) they fail to extract the geometry priors from single-view images due to the lack of multi-view supervision. In this paper, we propose a Geometry-enhanced NeRF (G-NeRF), which seeks to enhance the geometry priors by a geometry-guided multi-view synthesis approach, followed by a depth-aware training. In the synthesis process, inspired that existing 3D GAN models can unconditionally synthesize high-fidelity multi-view images, we seek to adopt off-the-shelf 3D GAN models, such as EG3D, as a free source to provide geometry priors through synthesizing multi-view data. Simultaneously, to further improve the geometry quality of the synthetic data, we introduce a truncation method to effectively sample latent codes within 3D GAN models. To tackle the absence of multi-view supervision for single-view images, we design the depth-aware training approach, incorporating a depth-aware discriminator to guide geometry priors through depth maps. Experiments demonstrate the effectiveness of our method in terms of both qualitative and quantitative results.

Visual content creation has aroused a surge of interest given its applications in mobile photography and AR/VR. Portrait style transfer and 3D recovery from monocular images as two representative tasks have so far evolved independently. In this paper, we make a connection between the two, and tackle the challenging task of 3D portrait stylization – modeling high-fidelity 3D stylized avatars from captured 2D portrait images. However, naively combining the techniques from the two isolated areas may suffer from either inadequate stylization or absence of 3D assets. To this end, we propose 3DToonify, a new framework that introduces a progressive training scheme to achieve 3D style adaption on spatial neural representation (SNR). SNR is constructed with implicit fields and they are dynamically optimized by the progressive training scheme, which consists of three stages: guided prior learning, deformable geometry adaption and explicit texture adaption. In this way, stylized geometry and texture are learned in SNR in an explicit and structured way with only a single stylized exemplar needed. Moreover, our method obtains style-adaptive underlying structures (i.e., deformable geometry and exaggerated texture) and view-consistent stylized avatar rendering from arbitrary novel viewpoints. Both qualitative and quantitative experiments have been conducted to demonstrate the effectiveness and superiority of our method for automatically generating exemplar-guided 3D stylized avatars.

Multi-view depth estimation has achieved impressive performance over various benchmarks. However, almost all current multi-view systems rely on given ideal camera poses, which are unavailable in many real-world scenarios, such as autonomous driving. In this work, we propose a new robustness benchmark to evaluate the depth estimation system under various noisy pose settings. Surprisingly, we find current multi-view depth estimation methods or single-view and multi-view fusion methods will fail when given noisy pose settings. To address this challenge, we propose a single-view and multi-view fused depth estimation system, which adaptively integrates high-confident multi-view and single-view results for both robust and accurate depth estimations. The adaptive fusion module performs fusion by dynamically selecting high-confidence regions between two branches based on a wrapping confidence map. Thus, the system tends to choose the more reliable branch when facing textureless scenes, inaccurate calibration, dynamic objects, and other degradation or challenging conditions. Our method outperforms state-of-the-art multi-view and fusion methods under robustness testing. Furthermore, we achieve state-of-the-art performance on challenging benchmarks (KITTI and DDAD) when given accurate pose estimations.

In this work, we present a novel dense-correspondence method for 6DoF object pose estimation from a single RGB-D image. While many existing data-driven methods achieve impressive performance, they tend to be time-consuming due to their reliance on rendering-based refinement approaches. To circumvent this limitation, we present HiPose, which establishes 3D-3D correspondences in a coarse-to-fine manner with a hierarchical binary surface encoding. Unlike previous dense-correspondence methods, we estimate the correspondence surface by employing point-to-surface matching and iteratively constricting the surface until it becomes a correspondence point while gradually removing outliers. Extensive experiments on public benchmarks LM-O, YCB-V, and T-Less demonstrate that our method surpasses all refinement-free methods and is even on par with expensive refinement-based approaches. Crucially, our approach is computationally efficient and enables real-time critical applications with high accuracy requirements. Code and models will be released.

Reconstructing 3D hand mesh robustly from a single image is very challenging, due to the lack of diversity in existing real-world datasets. While data synthesis helps relieve the issue, the syn-to-real gap still hinders its usage. In this work, we present HandBooster, a new approach to uplift the data diversity and boost the 3D hand-mesh reconstruction performance by training a conditional generative space on hand-object interactions and purposely sampling the space to synthesize effective data samples. First, we construct versatile content-aware conditions to guide a diffusion model to produce realistic images with diverse hand appearances, poses, views, and backgrounds; favorably, accurate 3D annotations are obtained for free. Then, we design a novel condition creator based on our similarity-aware distribution sampling strategies to deliberately find novel and realistic interaction poses that are distinctive from the training set. When partnered with our approach, several recent methods can significantly improve their performance beyond the SOTA on the HO3D and DexYCB benchmarks. Our code and data will be released upon publication.

Text-driven 3D scene generation techniques have made rapid progress in recent years. Their success is mainly attributed to using existing generative models to iteratively perform image warping and inpainting to generate 3D scenes. However, these methods heavily rely on the outputs of existing models, leading to error accumulation in geometry and appearance that prevent the models from being used in various scenarios (e.g., outdoor and unreal scenarios). To address this limitation, we generatively refine the newly generated local views by querying and aggregating global 3D information, and then progressively generate the 3D scene. Specifically, we employ a tri-plane features-based NeRF as a unified representation of the 3D scene to constrain global 3D consistency, and propose a generative refinement network to synthesize new contents with higher quality by exploiting the natural image prior from 2D diffusion model as well as the global 3D information of the current scene. Our extensive experiments demonstrate that, in comparison to previous methods, our approach supports wide variety of scene generation and arbitrary camera trajectories with improved visual quality and 3D consistency. Codes will be released.

We propose NViST, a transformer-based model for novel-view synthesis from a single image, trained on a large-scale dataset of in-the-wild images with complex backgrounds.NViST transforms image inputs directly into a radiance field, adopting a scalable transformer-based architecture. In practice, NViST exploits the self-supervised features learnt by a masked autoencoder (MAE), and learns a novel decoder that translates features to 3D tokens via cross-attention and adaptive layer normalization.Our model is efficient at inference since only a single forward-pass is needed to predict a 3D representation, unlike methods that require test-time optimization or sampling multiple times such as 3D-aware diffusion models.We tackle further limitations of current new-view synthesis models. First, unlike most generative models that are trained in a category-specific manner, often on synthetic datasets or on masked inputs, our model is trained on MVImgNet, a large-scale dataset of real-world, casually-captured videos containing hundreds of object categories with diverse backgrounds. Secondly, our model does not require canonicalization of the training data – i.e. aligning all objects with a frontal view – only needing relative pose as conditioning at training time which removes a substantial barrier to it being used on casually captured datasets. We show results on unseen objects and categories on MVImgNet and even casual phone captures. We conduct qualitative and quantitative evaluations on MVImgNet and ShapeNet to show that our model represents a step forward towards enabling true in-the-wild novel-view synthesis from a single image.

The increased demand for 3D data in AR/VR, robotics and gaming applications, gave rise to powerful generative pipelines capable of synthesizing high-quality 3D objects. Most of these models rely on the Score Distillation Sampling (SDS) algorithm to optimize a 3D representation such that the rendered image maintains a high likelihood as evaluated by a pre-trained diffusion model. However, this distillation process involves finding a correct mode in the high-dimensional and large-variance distribution produced by the diffusion model. This task is challenging and often leads to issues such as over-saturation, over-smoothing, and Janus-like artifacts in the 3D generation. In this paper, we propose a novel learning paradigm for 3D synthesis that utilizes pre-trained diffusion models. Instead of focusing on mode-seeking, our method directly models the distribution discrepancy between multi-view renderings and diffusion priors in an adversarial manner, which unlocks the generation of high-fidelity and photorealistic 3D content, conditioned on a single image and prompt. Moreover, by harnessing the latent space of GANs and expressive diffusion model priors, our method enables a wide variety of 3D applications including single-view reconstruction, high diversity generation and continuous 3D interpolation in open domain. Our experiments demonstrate the superiority of our pipeline compared to previous works in terms of generation quality and diversity.

We introduce the Splatter Image, an ultra-efficient approach for monocular 3D object reconstruction.Splatter Image is based on Gaussian Splatting, which allows fast and high-quality reconstruction of 3D scenes from multiple images.We apply Gaussian Splatting to monocular reconstruction by learning a neural network that, at test time, performs reconstruction in a feed-forward manner, at 38 FPS.Our main innovation is the surprisingly straightforward design of this network, which, using 2D operators, maps the input image to one 3D Gaussian per pixel.The resulting set of Gaussians thus has the form an image, the Splatter Image.We further extend the method take several images as input via cross-view attention.Owning to the speed of the renderer (588 FPS), we use a single GPU for training while generating entire images at each iteration to optimize perceptual metrics like LPIPS.On several synthetic, real, multi-category and large-scale benchmark datasets, we achieve better results in terms of PSNR, LPIPS, and other metrics while training and evaluating much faster than prior works.Code, models and more results are available at https://szymanowiczs.github.io/splatter-image.

Human-object contact serves as a strong cue to understand how humans physically interact with objects. Nevertheless, it is not widely explored to utilize human-object contact information for the joint reconstruction of 3D human and object from a single image. In this work, we present a novel joint 3D human-object reconstruction method (CONTHO) that effectively exploits contact information between humans and objects. There are two core designs in our system: 1) 3D-guided contact estimation and 2) contact-based 3D human and object refinement. First, for accurate human-object contact estimation, CONTHO initially reconstructs 3D humans and objects and utilizes them as explicit 3D guidance for contact estimation. Second, to refine the initial reconstructions of 3D human and object, we propose a novel contact-based refinement Transformer that effectively aggregates human features and object features based on the estimated human-object contact. The proposed contact-based refinement prevents the learning of erroneous correlation between human and object, which enables accurate 3D reconstruction. As a result, our CONTHO achieves state-of-the-art performance in both human-object contact estimation and joint reconstruction of 3D human and object. The codes are available in https://github.com/dqj5182/CONTHO_RELEASE.

Recent works on text-to-3d generation show that using only 2D diffusion supervision for 3D generation tends to produce results with inconsistent appearances (e.g., faces on the back view) and inaccurate shapes (e.g., animals with extra legs). Existing methods mainly address this issue by retraining diffusion models with images rendered from 3D data to ensure multi-view consistency while struggling to balance 2D generation quality with 3D consistency. In this paper, we present a new framework Sculpt3D that equips the current pipeline with explicit injection of 3D priors from retrieved reference objects without re-training the 2D diffusion model. Specifically, we demonstrate that high-quality and diverse 3D geometry can be guaranteed by keypoints supervision through a sparse ray sampling approach. Moreover, to ensure accurate appearances of different views, we further modulate the output of the 2D diffusion model to the correct patterns of the template views without altering the generated object's style. These two decoupled designs effectively harness 3D information from reference objects to generate 3D objects while preserving the generation quality of the 2D diffusion model. Extensive experiments show our method can largely improve the multi-view consistency while retaining fidelity and diversity.

Estimating the pose of objects from images is a crucial task of 3D scene understanding, and recent approaches have shown promising results on very large benchmarks. However, these methods experience a significant performance drop when dealing with unseen objects. We believe that it results from the limited generalizability of image features. To address this problem, we have an in-depth analysis on the features of diffusion models, e.g. Stable Diffusion, which hold substantial potential for modeling unseen objects. Based on this analysis, we then innovatively introduce these diffusion features for object pose estimation. To achieve this, we propose three distinct architectures that can effectively capture and aggregate diffusion features of different granularity, greatly improving the generalizability of object pose estimation. Our approach outperforms the state-of-the-art methods by a considerable margin on three popular benchmark datasets, LM, O-LM, and T-LESS. In particular, our method achieves higher accuracy than the previous best arts on unseen objects: 98.2% vs. 93.5% on Unseen LM, 85.9% vs. 76.3% on Unseen O-LM, showing the strong generalizability of our method.

Monocular 3D object detection has attracted widespread attention due to its potential to accurately obtain object 3D localization from a single image at a low cost. Depth estimation is an essential but challenging subtask of monocular 3D object detection due to the ill-posedness of 2D to 3D mapping. Many methods explore multiple local depth clues such as object heights and keypoints and then formulate the object depth estimation as an ensemble of multiple depth predictions to mitigate the insufficiency of single-depth information. However, the errors of existing multiple depths tend to have the same sign, which hinders them from neutralizing each other and limits the overall accuracy of combined depth. To alleviate this problem, we propose to increase the complementarity of depths with two novel designs. First, we add a new depth prediction branch named complementary depth that utilizes global and efficient depth clues from the entire image rather than the local clues to reduce the correlation of depth predictions. Second, we propose to fully exploit the geometric relations between multiple depth clues to achieve complementarity in form. Benefiting from these designs, our method achieves higher complementarity. Experiments on the KITTI benchmark demonstrate that our method achieves state-of-the-art performance without introducing extra data. In addition, complementary depth can also be a lightweight and plug-and-play module to boost multiple existing monocular 3d object detectors. Code is available at https://github.com/elvintanhust/MonoCD.

We introduce MultiDiff, a novel approach for consistent novel view synthesis of scenes from a single RGB image. The task of synthesising novel views from a single reference image is highly ill-posed by nature, as there exist multiple, plausible explanations for unobserved areas.To address this issue, we incorporate strong priors in form of monocular depth predictors and video-diffusion models. Monocular depth enables us to condition our model on warped reference images for the target views, increasing geometric stability. The video-diffusion prior provides a strong proxy for 3D scenes, allowing the model to learn continuous and pixel-accurate correspondences across generated images.In contrast to approaches relying on autoregressive image generation that are prone to drifts and error accumulation, MultiDiff jointly synthesizes a sequence of frames yielding high-quality and multi-view consistent results -- even for long-term scene generation with large camera movements, while reducing inference time by an order of magnitude.For additional consistency and image quality improvements, we introduce a novel, structured noise distribution. Our experimental results demonstrate that MultiDiff outperforms state-of-the-art methods on the challenging, real-world datasets RealEstate10K and ScanNet.Finally, our model naturally supports multi-view consistent editing without the need for further tuning.

Monocular 3D detectors achieve remarkable performance on cars and smaller objects. However, their performance drops on larger objects, leading to fatal accidents. Some attribute the failures to training data scarcity or their receptive field requirements of large objects. In this paper, we highlight this understudied problem of generalization to large objects and find that modern frontal detectors struggle to generalize to large objects even on balanced datasets. We argue that the cause of failure is the sensitivity of depth regression losses to noise of larger objects. To bridge this gap, we comprehensively investigate regression and dice losses, examining their robustness under varying error levels and object sizes. We mathematically prove that the dice loss leads to superior noise-robustness and model convergence for large objects compared to regression losses for a simplified case. Leveraging our theoretical insights, we propose SeaBird (Segmentation in Bird’s View) as the first step towards generalizing to large objects. SeaBird effectively integrates BEV segmentation on foreground objects for 3D detection, with the segmentation head trained with the dice loss. SeaBird achieves SoTA results on the KITTI-360 leaderboard and improves existing detectors on nuScenes leaderboard, particularly for large objects. Our code and models will be publicly available.

Monocular 3D detection is a challenging task due to the lack of accurate 3D information. Existing approaches typically rely on geometry constraints and dense depth estimates to facilitate the learning, but often fail to fully exploit the benefits of three-dimensional feature extraction in frustum and 3D space. In this paper, we propose OccupancyM3D, a method of learning occupancy for monocular 3D detection. It directly learns occupancy in frustum and 3D space, leading to more discriminative and informative 3D features and representations. Specifically, by using synchronized raw sparse LiDAR point clouds, we define the space status and generate voxel-based occupancy labels. We formulate occupancy prediction as a simple classification problem and design associated occupancy losses. Resulting occupancy estimates are employed to enhance original frustum/3D features. As a result, experiments on KITTI and Waymo open datasets demonstrate that the proposed method achieves a new state of the art and surpasses other methods by a significant margin.

We present a method for automatically modifying a NeRF representation based on a single observation of a non-rigid transformed version of the original scene.Our method defines the transformation as a 3D flow,specifically as a weighted linear blending of rigid transformations of 3D anchor points that are defined on the surface of the scene.In order to identify anchor points, we introduce a novel correspondence algorithm that first matches RGB-based pairs,then leverages multi-view information and 3D reprojection to robustly filter false positives in two steps. We also introduce a new dataset for exploring the problem of modifying a NeRF scene through a single observation. Our dataset contains 113 scenes leveraging 47 3D assets. We show that our proposed method outperforms NeRF editing methods as well as diffusion-based methods,and we also explore different methods for filtering correspondences.

Recently, the authors of Zero-1-to-3 demonstrated that a latent diffusion model, pretrained with Internet-scale data, can not only address the single-view 3D object reconstruction task but can even attain SOTA results in it. However, when applied to the task of single-view 3D clothed human reconstruction, Zero-1-to-3 (and related models) are unable to compete with the corresponding SOTA methods in this field despite being trained on clothed human data.In this work, we aim to tailor Zero-1-to-3’s approach to the single-view 3D clothed human reconstruction task in a much more principled and structured manner. To this end, we propose R-Cyclic Diffuser, a framework that adapts Zero-1-to-3’s novel approach to clothed human data by fusing it with a pixel-aligned implicit model.R-Cyclic Diffuser offers a total of three new contributions. The first and primary contribution is R-Cyclic Diffuser’s cyclical conditioning mechanism for novel view synthesis. This mechanism directly addresses the view inconsistency problem faced by Zero-1-to-3 and related models. Secondly, we further enhance this mechanism with two key features - Lateral Inversion Constraint and Cyclic Noise Selection. Both features are designed to regularize and restrict the randomness of outputs generated by a latent diffusion model. Thirdly, we show how SMPL-X body priors can be incorporated in a latent diffusion model such that novel views of clothed human bodies can be generated much more accurately. Our experiments show that R-Cyclic Diffuser is able to outperform current SOTA methods in single-view 3D clothed human reconstruction both qualitatively and quantitatively. Our code is made publicly available at https://github.com/kcyt/r-cyclic-diffuser.

Semantic scene completion (SSC) aims to predict complete 3D voxel occupancy and semantics from a single-view RGB-D image, and recent SSC methods commonly adopt multi-modal inputs. However, our investigation reveals two limitations: ineffective feature learning from single modalities and overfitting to limited datasets. To address these issues, this paper proposes a novel SSC framework - Potential Unleashing Network (PUNet) - with a fresh perspective of optimizing gradient updates. The proposed PUNet introduces two core modules: a cross-modal modulation enabling the interdependence of gradient flows between modalities, and a customized adversarial training scheme leveraging dynamic gradient competition. Specifically, the cross-modal modulation adaptively re-calibrates the features to better excite representation potentials from each single modality. The adversarial training employs a minimax game of evolving gradients, with customized guidance to strengthen the generator's perception of visual fidelity from both geometric completeness and semantic correctness. Extensive experimental results demonstrate that PUNet outperforms state-of-the-art SSC methods by a large margin, providing a promising direction for improving the effectiveness and generalization of SSC methods. Our code is in the Appendix.

Recent advancements in 3D reconstruction from single images have been driven by the evolution of generative models. Prominent among these are methods based on Score Distillation Sampling (SDS) and the adaptation of diffusion models in the 3D domain. Despite their progress, these techniques often face limitations due to slow optimization or rendering processes, leading to extensive training and optimization times. In this paper, we introduce a novel approach for single-view reconstruction that efficiently generates a 3D model from a single image via feed-forward inference. Our method utilizes two transformer-based networks, namely a point decoder and a triplane decoder, to reconstruct 3D objects using a hybrid Triplane-Gaussian intermediate representation. This hybrid representation strikes a balance, achieving a faster rendering speed compared to implicit representations while simultaneously delivering superior rendering quality than explicit representations. The point decoder is designed for generating point clouds from single images, offering an explicit representation which is then utilized by the triplane decoder to query Gaussian features for each point. This design choice addresses the challenges associated with directly regressing explicit 3D Gaussian attributes characterized by their non-structural nature. Subsequently, the 3D Gaussians are decoded by an MLP to enable rapid rendering through splatting. Both decoders are built upon a scalable, transformer-based architecture and have been efficiently trained on large-scale 3D datasets. The evaluations conducted on both synthetic datasets and real-world images demonstrate that our method not only achieves higher quality but also ensures a faster runtime in comparison to previous state-of-the-art techniques.

We present a 3D-aware one-shot head reenactment method based on a fully volumetric neural disentanglement framework for source appearance and driver expressions. Our method is real-time and produces high-fidelity and view-consistent output, suitable for 3D teleconferencing systems based on holographic displays. Existing cutting-edge 3D-aware reenactment methods often use neural radiance fields or 3D meshes to produce view-consistent appearance encoding, but, at the same time, they rely on linear face models, such as 3DMM, to achieve its disentanglement with facial expressions. As a result, their reenactment results often exhibit identity leakage from the driver or have unnatural expressions. To address these problems, we propose a neural self-supervised disentanglement approach that lifts both the source image and driver video frame into a shared 3D volumetric representation based on tri-planes. This representation can then be freely manipulated with expression tri-planes extracted from the driving images and rendered from an arbitrary view using neural radiance fields. We achieve this disentanglement via self-supervised learning on a large in-the-wild video dataset. We further introduce a highly effective fine-tuning approach to improve the generalizability of the 3D lifting using the same real-world data. We demonstrate state-of-the-art performance on a wide range of datasets, and also showcase high-quality 3D-aware head reenactment on highly challenging and diverse subjects, including non-frontal head poses and complex expressions for both source and driver.

Recently, high-fidelity scene reconstruction with an optimized 3D Gaussian splat representation has been introduced for novel view synthesis from sparse image sets. Making such representations suitable for applications like network streaming and rendering on low-power devices requires significantly reduced memory consumption as well as improved rendering efficiency.We propose a compressed 3D Gaussian splat representation that utilizes sensitivity-aware vector clustering with quantization-aware training to compress directional colors and Gaussian parameters. The learned codebooks have low bitrates and achieve a compression rate of up to $31\times$ on real-world scenes with only minimal degradation of visual quality.We demonstrate that the compressed splat representation can be efficiently rendered with hardware rasterization on lightweight GPUs at up to $4\times$ higher framerates than reported via an optimized GPU compute pipeline.Extensive experiments across multiple datasets demonstrate the robustness and rendering speed of the proposed approach.

Recent advances in generative diffusion models have enabled the previously unfeasible capability of generating 3D assets from a single input image or a text prompt. In this work, we aim to enhance the quality and functionality of these models for the task of creating controllable, photorealistic human avatars. We achieve this by integrating a 3D morphable model into the state-of-the-art multiview-consistent diffusion approach. First, we demonstrate that proper conditioning of a generative pipeline with the articulated 3D model enhances the baseline model performance on the task of novel view synthesis from a single image. Next, we introduce a training regime that enables the animation of the reconstructed model with new facial expressions and body poses. To the best of our knowledge, our proposed model is the first to enable the creation of a 3D-consistent, animatable and photorealistic human avatar from a single image of an unseen subject. The code for our project will be made publicly available.

This work presents Depth Anything, a highly practical solution for robust monocular depth estimation. Without pursuing novel technical modules, we aim to build a simple yet powerful foundation model dealing with any images under any circumstances. To this end, we scale up the dataset by designing a data engine to collect and automatically annotate large-scale unlabeled data (~17M), which significantly enlarges the data coverage and thus is able to reduce the generalization error. However, naively utilizing pseudo labels in training suffers severe performance drop. We investigate two simple yet effective strategies that make data scaling-up promising. First, a more challenging optimization target is created by leveraging data augmentation tools. It compels the model to actively seek extra visual knowledge and acquire more robust representations. Second, an auxiliary supervision is developed to enforce the model to inherit rich semantic priors from pre-trained encoders. We evaluate its zero-shot capabilities in various situations, including six public datasets and randomly captured photos. It demonstrates impressive generalization ability. Through fine-tuning it on NYUv2 and KITTI, new SOTAs are set. The models, training and evaluation code will all be released.

We introduce SAOR, a novel approach for estimating the 3D shape, texture, and viewpoint of an articulated object from a single image captured in the wild. Unlike prior approaches that rely on pre-defined category-specific 3D templates or tailored 3D skeletons, SAOR learns to articulate shapes from single-view image collections with a skeleton-free part-based model without requiring any 3D object shape priors. To prevent ill-posed solutions, we propose a cross-instance consistency loss that exploits disentangled object shape deformation and articulation. This is helped by a new silhouette-based sampling mechanism to enhance viewpoint diversity during training. Our method only requires estimated object silhouettes and relative depth maps from off-the-shelf pre-trained networks during training. At inference time, given a single-view image, it efficiently outputs an explicit mesh representation. We obtain improved qualitative and quantitative results on challenging quadruped animals compared to relevant existing work.

We introduce EscherNet, a multi-view conditioned diffusion model for view synthesis. EscherNet learns implicit and generative 3D representations coupled with a specialised camera positional encoding, allowing precise and continuous relative control of the camera transformation between an arbitrary number of reference and target views. EscherNet offers exceptional generality, flexibility, and scalability in view synthesis --- it can generate more than 100 consistent target views simultaneously on a single consumer-grade GPU, despite being trained with a fixed number of 3 reference views to 3 target views. As a result, EscherNet not only addresses zero-shot novel view synthesis, but also naturally unifies single- and multi-image 3D reconstruction, combining these diverse tasks into a single, cohesive framework. Our extensive experiments demonstrate that EscherNet achieves state-of-the-art performance in multiple benchmarks, even when compared to methods specifically tailored for each individual problem. This remarkable versatility opens up new directions for designing scalable neural architectures for 3D vision. Project page: https://kxhit.github.io/EscherNet.

Human hands are highly articulated and versatile at handling objects. Jointly estimating the 3D poses of a hand and the object it manipulates from a monocular camera is challenging due to frequent occlusions. Thus, existing methods often rely on intermediate 3D shape representations to increase performance. These representations are typically explicit, such as 3D point clouds or meshes, and thus provide information in the direct surroundings of the intermediate hand pose estimate. To address this, we introduce HOISDF, a Signed Distance Field (SDF) guided hand-object pose estimation network, which jointly exploits hand and object SDFs to provide a global, implicit representation over the complete reconstruction volume. Specifically, the role of the SDFs is threefold: equip the visual encoder with implicit shape information, help to encode hand-object interactions, and guide the hand and object pose regression via SDF-based sampling and by augmenting the feature representations. We show that HOISDF achieves state-of-the-art results on hand-object pose estimation benchmarks (DexYCB and HO3Dv2).

We present a new multi-modal face image generation method that converts a text prompt and a visual input, such as a semantic mask or scribble map, into a photo-realistic face image. To do this, we combine the strengths of Generative Adversarial networks (GANs) and diffusion models (DMs) by employing the multi-modal features in the DM into the latent space of the pre-trained GANs. We present a simple mapping and a style modulation network to link two models and convert meaningful representations in feature maps and attention maps into latent codes. With GAN inversion, the estimated latent codes can be used to generate 2D or 3D-aware facial images. We further present a multi-step training strategy that reflects textual and structural representations into the generated image. Our proposed network produces realistic 2D, multi-view, and stylized face images, which align well with inputs. We validate our method by using pre-trained 2D and 3D GANs, and our results outperform existing methods. Our project page is available at https://github.com/1211sh/Diffusiondriven_GAN-Inversion/.

In this paper, we firstly consider view-dependent effects into single image-based novel view synthesis (NVS) problems. For this, we propose to exploit the camera motion priors in NVS to model view-dependent appearance or effects (VDE) as the negative disparity in the scene. By recognizing specularities \enquote{follow} the camera motion, we infuse VDEs into the input images by aggregating input pixel colors along the negative depth region of the epipolar lines. Also, we propose a `relaxed volumetric rendering' approximation that allows computing the densities in a single pass, improving efficiency for NVS from single images. Our method can learn single-image NVS from image sequences only, which is a completely self-supervised learning method, for the first time requiring neither depth nor camera pose annotations.We present extensive experiment results and show that our proposed method can learn NVS with VDEs, outperforming the SOTA single-view NVS methods on the RealEstate10k and MannequinChallenge datasets.

Generating vivid and emotional 3D co-speech gestures is crucial for virtual avatar animation in human-machine interaction applications. While the existing methods enable generating the gestures to follow a single emotion label, they overlook that long gesture sequence modeling with emotion transition is more practical in real scenes. In addition, the lack of large-scale available datasets with emotional transition speech and corresponding 3D human gestures also limits the addressing of this task. To fulfill this goal, we first incorporate the ChatGPT-4 and an audio inpainting approach to construct the high-fidelity emotion transition human speeches. Considering obtaining the realistic 3D pose annotations corresponding to the dynamically inpainted emotion transition audio is extremely difficult, we propose a novel weakly supervised training strategy to encourage authority gesture transitions. Specifically, to enhance the coordination of transition gestures w.r.t. different emotional ones, we model the temporal association representation between two different emotional gesture sequences as style guidance and infuse it into the transition generation. We further devise an emotion mixture mechanism that provides weak supervision based on a learnable mixed emotion label for transition gestures. Last, we present a keyframe sampler to supply effective initial posture cues in long sequences, enabling us to generate diverse gestures. Extensive experiments demonstrate that our method outperforms the state-of-the-art models constructed by adapting single emotion-conditioned counterparts on our newly defined emotion transition task and datasets.

State-of-the-art single-view 360-degree room layout reconstruction methods formulate the problem as a high-level 1D (per-column) regression task. On the other hand, traditional low-level 2D layout segmentation is simpler to learn and can represent occluded regions, but it requires complex post-processing for the targeting layout polygon and sacrifices accuracy. We present Seg2Reg to render 1D layout depth regression from the 2D segmentation map in a differentiable and occlusion-aware way, marrying the merits of both sides.Specifically, our model predicts floor-plan density for the input equirectangular 360-degree image. Formulating the 2D layout representation as a density field enables us to employ `flattened' volume rendering to form 1D layout depth regression. In addition, we propose a novel 3D warping augmentation on layout to improve generalization. Finally, we re-implement recent room layout reconstruction methods into our codebase for benchmarking and explore modern backbones and training techniques to serve as the strong baseline. Our model significantly outperforms previous arts. The code will be made available upon publication.

This paper focuses on self-supervised monocular depth estimation in dynamic scenes trained on monocular videos. Existing methods jointly estimate pixel-wise depth and motion, relying mainly on an image reconstruction loss. Dynamic regions remain a critical challenge for these methods due to the inherent ambiguity in depth and motion estimation, resulting in inaccurate depth estimation. This paper proposes a self-supervised training framework exploiting pseudo depth labels for dynamic regions from training data. The key contribution of our framework is to decouple depth estimation for static and dynamic regions of images in the training data. We start with an unsupervised depth estimation approach, which provides reliable depth estimates for static regions and motion cues for dynamic regions and allows us to extract moving object information at the instance level. In the next stage, we use an object network to estimate the depth of those moving objects assuming rigid motions. Then, we propose a new scale alignment module to address the scale ambiguity between estimated depths for static and dynamic regions. We can then use the depth labels generated to train an end-to-end depth estimation network and improve its performance. Extensive experiments on the Cityscapes and KITTI datasets show that our self-training strategy consistently outperforms existing self-/unsupervised depth estimation methods.

Current methods for 2D and 3D object understanding struggle with severe occlusions in busy urban environments, partly due to the lack of large-scale labeled ground-truth annotations for learning occlusion. In this work, we introduce a novel framework for automatically generating a large, realistic dataset of dynamic objects under occlusions using freely available time-lapse imagery. By leveraging off-the-shelf 2D (bounding box, segmentation, keypoint) and 3D (pose, shape) predictions as pseudo-groundtruth, unoccluded 3D objects are identified automatically and composited into the background in a clip-art style, ensuring realistic appearances and physically accurate occlusion configurations. The resulting clip-art image with pseudo-groundtruth enables efficient training of object reconstruction methods that are robust to occlusions. Our method demonstrates significant improvements in both 2D and 3D reconstruction, particularly in scenarios with heavily occluded objects like vehicles and people in urban scenes.

We present $DiffPortrait3D$, a conditional diffusion model that is capable of synthesizing 3D-consistent photo-realistic novel views from as few as a single in-the-wild portrait. Specifically, given a single RGB input, we aim to synthesize plausible but consistent facial details rendered from novel camera views with retained both identity and facial expression. In lieu of time-consuming optimization and fine-tuning, our zero-shot method generalizes well to arbitrary face portraits with unposed camera views, extreme facial expressions, and diverse artistic depictions. At its core, we leverage the generative prior of 2D diffusion models pre-trained on large-scale image datasets as our rendering backbone, while the denoising is guided with disentangled attentive control of appearance and camera pose. To achieve this, we first inject the appearance context from the reference image into the self-attention layers of the frozen UNets. The rendering view is then manipulated with a novel conditional control module that interprets the camera pose by watching a condition image of a crossed subject from the same view. Furthermore, we insert a trainable cross-view attention module to enhance view consistency, which is further strengthened with a novel 3D-aware noise generation process during inference. We demonstrate state-of-the-art results both qualitatively and quantitatively on our challenging in-the-wild and multi-view benchmarks.

The lifting of a 3D structure and camera from 2D landmarks is at the cornerstone of the discipline of computer vision. Traditional methods have been confined to specific rigid objects, such as those in Perspective-n-Point (PnP) problems, but deep learning has expanded our capability to reconstruct a wide range of object classes (e.g. C3DPO and PAUL) with resilience to noise, occlusions, and perspective distortions. However, all these techniques have been limited by the fundamental need to establish correspondences across the 3D training data, significantly limiting their utility to applications where one has an abundance of ``in-correspondence'' 3D data. Our approach harnesses the inherent permutation equivariance of transformers to manage varying numbers of points per 3D data instance, withstands occlusions, and generalizes to unseen categories. We demonstrate state-of-the-art performance across 2D-3D lifting task benchmarks. Since our approach can be trained across such a broad class of structures, we refer to it simply as a 3D Lifting Foundation Model (3D-LFM) -- the first of its kind.

We propose a single-shot approach to determining 6 DoF pose of an object with available 3D CAD model from a single RGB image. Our method, dubbed MRC-Net, comprises two stages. The first performs pose classification and renders the 3D object in the classified pose. The second stage performs regression to predict fine-grained relative pose within class. Connecting the two stages is a novel multi-scale residual correlation (MRC) layer that captures high-and-low level correspondences between the input image and rendering from first stage. MRC-Net employs a Siamese network with shared weights between both stages to learn embeddings for input and rendered images. To mitigate ambiguity when predicting discrete pose class labels on symmetric objects, we use soft probabilistic labels to define pose class in the first stage. We demonstrate state-of-the-art accuracy, outperforming all competing RGB-based methods on four challenging BOP benchmark datasets: T-LESS, LM-O, YCB-V, and ITODD. Our method is non-iterative and requires no complex post-processing.

Text-guided domain adaptation and generation of 3D-aware portraits find many applications in various fields. However, due to the lack of training data and the challenges in handling the high variety of geometry and appearance, the existing methods for these tasks suffer from issues like inflexibility, instability, and low fidelity. In this paper, we propose a novel framework DiffusionGAN3D, which boosts text-guided 3D domain adaptation and generation by combining 3D GANs and diffusion priors. Specifically, we integrate the pre-trained 3D generative models (e.g., EG3D) and text-to-image diffusion models. The former provides a strong foundation for stable and high-quality avatar generation from text. And the diffusion models in turn offer powerful priors and guide the 3D generator finetuning with informative direction to achieve flexible and efficient text-guided domain adaptation. To enhance the diversity in domain adaptation and the generation capability in text-to-avatar, we introduce the relative distance loss and case-specific learnable triplane respectively. Besides, we design a progressive texture refinement module to improve the texture quality for both tasks above. Extensive experiments demonstrate that the proposed framework achieves excellent results in both domain adaptation and text-to-avatar tasks, outperforming existing methods in terms of generation quality and efficiency. The project homepage is at https://younglbw.github.io/DiffusionGAN3D-homepage/.

Various applications require high-fidelity and artifact-free 3D human reconstructions. However, current implicit function-based methods inevitably produce artifacts while existing deformation methods are difficult to reconstruct high-fidelity humans wearing loose clothing. In this paper, we propose a two-stage deformation method named Vertex Shift (VS) for reconstructing clothed 3D humans from single images. Specifically, VS first stretches the estimated SMPL-X mesh into a coarse 3D human model using shift fields inferred from normal maps, then refines the coarse 3D human model into a detailed 3D human model via a graph convolutional network embedded with implicit-function-learned features. This ``stretch-refine'' strategy addresses both large deformations required for reconstructing loose clothing and delicate deformations for recovering intricate and detailed surfaces, achieving high-fidelity reconstructions that faithfully convey the pose, clothing, and surface details from the input images. The graph convolutional network's ability to exploit neighborhood vertices coupled with the advantages inherited from the deformation methods ensure VS rarely produces artifacts like distortions and non-human shapes and never produces artifacts like holes, broken parts, and dismembered limbs. As a result, VS can reconstruct high-fidelity and artifact-less clothed 3D humans from single images, even under scenarios of challenging poses and loose clothing. Experimental results on three benchmarks and two in-the-wild datasets demonstrate that VS significantly outperforms current state-of-the-art methods. The code and models of VS will be released.

Monocular 3D detection (M3D) aims for precise 3D object localization from a single-view image which usually involves labor-intensive annotation of 3D detection boxes. Weakly supervised M3D has recently been studied to obviate the 3D annotation process by leveraging many existing 2D annotations, but it often requires extra training data such as LiDAR point clouds or multi-view images which greatly degrades its applicability and usability in various applications. We propose SKD-WM3D, a weakly supervised monocular 3D detection framework that exploits depth information to achieve M3D with a single-view image exclusively without any 3D annotations or other training data. One key design in SKD-WM3D is a self-knowledge distillation framework, which transforms image features into 3D-like representations by fusing depth information and effectively mitigates the inherent depth ambiguity in monocular scenarios with little computational overhead in inference. In addition, we design an uncertainty-aware distillation loss and a gradient-targeted transfer modulation strategy which facilitate knowledge acquisition and knowledge transfer, respectively. Extensive experiments show that SKD-WM3D surpasses the state-of-the-art clearly and is even on par with many fully supervised methods.

Self-supervised monocular depth estimation (DE) is an approach to learning depth without costly depth ground truths. However, it often struggles with moving objects that violate the static scene assumption during training. To address this issue, we introduce a coarse-to-fine training strategy leveraging the ground contacting prior based on the observation that most moving objects in outdoor scenes contact the ground. In the coarse training stage, we exclude the objects in dynamic classes from the reprojection loss calculation to avoid inaccurate depth learning. To provide precise supervision on the depth of the objects, we present a novel Ground-contacting-prior Disparity Smoothness Loss (GDS-Loss) that encourages a DE network to align the depth of the objects with their ground-contacting points. Subsequently, in the fine training stage, we refine the DE network to learn the detailed depth of the objects from the reprojection loss, while ensuring accurate DE on the moving object regions by employing our regularization loss with a cost-volume-based weighting factor. Our overall coarse-to-fine training strategy can easily be integrated with existing DE methods without any modifications, significantly enhancing DE performance on challenging Cityscapes and KITTI datasets, especially in the moving object regions.

Reconstructing outdoor 3D scenes from temporal observations is a challenge that recent work on neural fields has offered a new avenue for. However, existing methods that recover scene properties, such as geometry, appearance, or radiance, solely from RGB captures often fail when handling poorly-lit or texture-deficient regions. Similarly, recovering scenes with scanning lidar sensors is also difficult due to their low angular sampling rate which makes recovering expansive real-world scenes difficult. Tackling these gaps, we introduce Gated Fields – a neural scene reconstruction method that utilizes active gated video sequences. To this end, we propose a neural rendering approach that seamlessly incorporates time-gated capture and illumination. Our method exploits the intrinsic depth cues in the gated videos, achieving precise and dense geometry reconstruction irrespective of ambient illumination conditions. We validate the method across day and night scenarios and find that Gated Fields compares favorably to RGB and LiDAR reconstruction methods

In this paper, we explore the potential of Snapshot Com- pressive Imaging (SCI) technique for recovering the under- lying 3D scene representation from a single temporal com- pressed image. SCI is a cost-effective method that enables the recording of high-dimensional data, such as hyperspec- tral or temporal information, into a single image using low- cost 2D imaging sensors. To achieve this, a series of spe- cially designed 2D masks are usually employed, which not only reduces storage requirements but also offers potential privacy protection. Inspired by this, to take one step further, our approach builds upon the powerful 3D scene represen- tation capabilities of neural radiance fields (NeRF). Specif- ically, we formulate the physical imaging process of SCI as part of the training of NeRF, allowing us to exploit its impressive performance in capturing complex scene struc- tures. To assess the effectiveness of our method, we con- duct extensive evaluations using both synthetic data and real data captured by our SCI system. Extensive experi- mental results demonstrate that our proposed approach sur- passes the state-of-the-art methods in terms of image re- construction and novel view image synthesis. Moreover, our method also exhibits the ability to restore high frame- rate multi-view consistent images by leveraging SCI and the rendering capabilities of NeRF. The code is available at https://github.com/WU-CVGL/SCINeRF.

Fourier occupancy field-based human reconstruction is a simple method that transforms the occupancy function of the 3D model into a multichannel 2D vector field. However, accurately estimating high-frequency information of the FOF is challenging, leading to geometric distortion and discontinuity. To this end, we propose a wavelet-based diffusion model to predict the FOF, extracting more high-frequency information and enhancing geometric stability. Our method comprises two interconnected tasks: texture estimation and geometry prediction. Initially, we predict the back-side texture from the input image, incorporating a style consistency constraint between the predicted back-side image and the original input image. To enhance network training effectiveness, we adopt a Siamese network training strategy. We introduce a wavelet-based diffusion model for geometric estimation to generate the Fourier occupancy field. First, we utilize an image encoder module to extract the features of the two images as conditions. Subsequently, we employ a conditional diffusion model to estimate the Fourier occupancy field in the wavelet domain. The predicted wavelet coefficients are then converted into the Fourier occupancy field using the inverse wavelet transform (IWT). A refinement network refines the predicted Fourier occupancy field with image features as guidance, yielding the final output. Through both quantitative and qualitative experiments, we demonstrate the state-of-the-art performance of our method in reconstructing single-view clothed human subjects.

A simple yet effective method for occlusion-robust 3D human mesh reconstruction from a single image is presented in this paper.Although many recent studies have shown the remarkable improvement in human mesh reconstruction, it is still difficult to generate accurate meshes when person-to-person occlusion occurs due to the ambiguity of who a body part belongs to.To address this problem, we propose an instance-aware contrastive learning scheme.Specifically, joint features belonging to the target human are trained to be proximate with the anchor feature (i.e., feature extracted from the body center position).On the other hand, anchor features of different human instances are forced to be far apart so that joint features of each person can be clearly distinguished from others.By interpreting the joint possession based on such contrastive learning scheme, the proposed method easily understands the spatial occupancy of body parts for each person in a given image, thus can reconstruct reliable human meshes even with severely overlapped cases between multiple persons.Experimental results on benchmark datasets demonstrate the robustness of the proposed method compared to previous approaches under person-to-person occlusions.The code and model are publicly available at: https://github.com/DCVL-3D/InstanceHMR_release.

We present a method for visual SLAM that can generalize to unseen scenes without the need for retraining and enjoys fast optimization. Existing methods struggle to generalize to novel scenes, i.e., they are optimized on a per-scene basis. Recently, neural scene representations have shown promise in SLAM to produce dense 3D reconstruction with high quality, at the cost of long training time. To overcome the limitations on generalization and efficiency, we propose IBD-SLAM, an Image-Based Depth fusion framework for generalizable SLAM. In particular, we adopt a Neural Radiance Field (NeRF) for scene representation. Inspired by image-based rendering, instead of learning a fixed grid of scene representation, we propose to learn image-based depth fusion, by deriving xyz-maps from the depth maps inferred from the given images. Once trained, the model can be applied to new uncalibrated monocular RGBD videos of unseen scenes, without the need for retraining. For any given new scene, only the pose parameters need to be optimized, which is very efficient. We thoroughly evaluate IBD-SLAM on public visual SLAM benchmarks, outperforming the previous state-of-the-art while being 10 times faster.

Recent progress in single-image 3D generation highlights the importance of multi-view coherency, leveraging 3D priors from large-scale diffusion models pretrained on Internet-scale images. However, the aspect of novel-view diversity remains underexplored within the research landscape due to the ambiguity in converting a 2D image into 3D content, where numerous potential shapes can emerge. Here, we aim to address this research gap by simultaneously addressing both consistency and diversity. Yet, striking a balance between these two aspects poses a considerable challenge due to their inherent trade-offs. This work introduces HarmonyView, a simple yet effective diffusion sampling technique adept at decomposing two intricate aspects in single-image 3D generation: consistency and diversity. This approach paves the way for a more nuanced exploration of the two critical dimensions within the sampling process. Moreover, we propose a new evaluation metric based on CLIP image and text encoders to comprehensively assess the diversity of the generated views, which closely aligns with human evaluators' judgments. In experiments, HarmonyView achieves a harmonious balance, demonstrating a win-win scenario in both consistency and diversity.

3D face reconstruction aims at generating high-fidelity 3D face shapes and textures from single-view or multi-view images. However, current prevailing facial texture generation methods generally suffer from low-quality texture, identity information loss, and inadequate handling of occlusions. To solve these problems, we introduce an Identity-Conditioned Latent Diffusion Model for face UV-texture generation (UV-IDM) to generate photo-realistic textures based on the Basel Face Model (BFM). UV-IDM leverages the powerful texture generation capacity of a latent diffusion model (LDM) to obtain detailed facial textures. To preserve the identity during the reconstruction procedure, we design an identity-conditioned module that can utilize any in-the-wild image as a robust condition for the LDM to guide texture generation. UV-IDM can be easily adapted to different BFM-based methods as a high-fidelity texture generator.Furthermore, in light of the limited accessibility of most existing UV-texture datasets, we build a large-scale and publicly available UV-texture dataset based on BFM, termed BFM-UV. Extensive experiments show that our UV-IDM can generate high-fidelity textures in 3D face reconstruction within seconds while maintaining image consistency, bringing new state-of-the-art performance in facial texture generation.

Editable 3D-aware generation, which supports user-interacted editing, has witnessed rapid development recently. However, existing editable 3D GANs either fail to achieve high-accuracy local editing or suffer from huge computational costs. We propose AttriHuman-3D, an editable 3D human generation model, which address the aforementioned problems with attribute decomposition and indexing. The core idea of the proposed model is to generate all attributes (e.g. human body, hair, clothes and so on) in an overall attribute space with six feature planes, which are then decomposed and manipulated with different attribute indexes. To precisely extract features of different attributes from the generated feature planes, we propose a novel attribute indexing method as well as an orthogonal projection regularization to enhance the disentanglement. We also introduce a hyper-latent training strategy and an attribute-specific sampling strategy to avoid style entanglement and misleading punishment from the discriminator. Our method allows users to interactively edit selected attributes in the generated 3D human avatars while keeping others fixed. Both qualitative and quantitative experiments demonstrate that our model provides a strong disentanglement between different attributes, allows fine-grained image editing and generates high-quality 3D human avatars.

Monocular Depth Estimation (MDE) is a fundamentalproblem in computer vision with numerous applications.Recently, LIDAR-supervised methods have achieved remarkable per-pixel depth accuracy in outdoor scenes. However, significant errors are typically found in the proximityof depth discontinuities, i.e., depth edges, which often hinder the performance of depth-dependent applications thatare sensitive to such inaccuracies, e.g., novel view synthesis and augmented reality. Since direct supervision for thelocation of depth edges is typically unavailable in sparseLIDAR-based scenes, encouraging the MDE model to produce correct depth edges is not straightforward. To the bestof our knowledge this paper is the first attempt to addressthe depth edges issue for LIDAR-supervised scenes. In thiswork we propose to learn to detect the location of depthedges from densely-supervised synthetic data, and use it togenerate supervision for the depth edges in the MDE training. To quantitatively evaluate our approach, and due tothe lack of depth edges GT in LIDAR-based scenes, wemanually annotated subsets of the KITTI and the DDADdatasets with depth edges ground truth. We demonstratesignificant gains in the accuracy of the depth edges withcomparable per-pixel depth accuracy on several challenging datasets. Code and datasets are available at https://github.com/liortalker/MindTheEdge.

3DiffTection introduces a novel method for 3D object detection from single images, utilizing a 3D-aware diffusion model for feature extraction. Addressing the resource-intensive nature of annotating large-scale 3D image data, our approach leverages pretrained diffusion models, traditionally used for 2D tasks, and adapts them for 3D detection through geometric and semantic tuning.Geometrically, we enhance the model to perform view synthesis from single images, incorporating an epipolar warp operator. This process utilizes easily accessible posed image data, eliminating the need for manual annotation. Semantically, the model is further refined on target detection data. Both stages utilize ControlNet, ensuring the preservation of original feature capabilities. Through our methodology, we obtain 3D-aware features that excel in identifying cross-view point correspondences. In 3D detection, 3DiffTection substantially surpasses previous benchmarks, \textit{e.g.,} Cube-RCNN, by 9.43% in AP3D on the Omni3D-ARkitscene dataset. Furthermore, 3DiffTection demonstrates robust label efficiency and generalizes well to cross-domain data, nearly matching fully-supervised models in zero-shot scenarios.

We present Bayesian Diffusion Models (BDM), a prediction algorithm that performs effective Bayesian inference by tightly coupling the top-down (prior) information with the bottom-up (data-driven) procedure via joint diffusion processes. We demonstrate the application of BDM on the 3D shape reconstruction task. Compared to standard deep learning data-driven approaches relying on supervised data, our BDM can bring in rich prior information trained in an unsupervised manner to improve the bottom-up 3D reconstruction. As opposed to the traditional Bayesian frameworks where explicitly learned prior and data-driven distributions are required for gradient computation and combination, BDM performs a seamless fusion of the two via coupled diffusion processes with learned gradient computation networks. The specialty of our Bayesian Diffusion Models (BDM) lies in its capability to engage the active and effective information exchange and fusion of the top-down and bottom-up processes where each itself is a diffusion process. We demonstrate state-of-the-art results on both synthetic and real-world benchmarks for 3D shape reconstruction.

Despite the growing demand for accurate surface normal estimation models, existing methods use general-purpose dense prediction models, adopting the same inductive biases as other tasks. In this paper, we discuss the inductive biases needed for surface normal estimation and propose to (1) utilize the per-pixel ray direction and (2) encode the relationship between neighboring surface normals by learning their relative rotation. The proposed method can generate crisp - yet, piecewise smooth - predictions for challenging in-the-wild images of arbitrary resolution and aspect ratio. Compared to a recent ViT-based state-of-the-art model, our method shows a stronger generalization ability, despite being trained on an orders of magnitude smaller dataset.

Monocular 3D lane detection has become a fundamental problem in the context of autonomous driving, which comprises the tasks of finding the road surface and locating lane markings. One major challenge lies in a flexible but robust line representation capable of modeling complex lane structures, while still avoiding unpredictable behavior. While previous methods rely on fully data-driven approaches, we instead introduce a novel approach LaneCPP that uses a continuous 3D lane detection model leveraging physical prior knowledge about the lane structure and road geometry. While our sophisticated lane model is capable of modeling complex road structures, it also shows robust behavior since physical constraints are incorporated by means of a regularization scheme that can be analytically applied to our parametric representation. Moreover, we incorporate prior knowledge about the road geometry into the 3D feature space by modeling geometry-aware spatial features, guiding the network to learn an internal road surface representation. In our experiments, we show the benefits of our contributions and prove the meaningfulness of using priors to make 3D lane detection more robust. The results show that LaneCPP achieves state-of-the-art performance in terms of F-Score and geometric errors.

Leveraging multi-view diffusion models as priors for 3D optimization have alleviated the problem of 3D consistency, e.g., the Janus face problem or the content drift problem, in zero-shot text-to-3D models.However, the 3D geometric fidelity of the output remains an unresolved issue; albeit the rendered 2D views are realistic, the underlying geometry may contain errors such as unreasonable concavities.In this work, we propose CorrespondentDream, an effective method to leverage annotation-free, cross-view correspondences yielded from the diffusion U-Net to provide additional 3D prior to the NeRF optimization process.We find that these correspondences are strongly consistent with human perception, and by adopting it in our loss design, we are able to produce NeRF models with geometries that are more coherent with common sense, e.g., more smoothed object surface, yielding higher 3D fidelity.We demonstrate the efficacy of our approach through various comparative qualitative results and a solid user study.

3D object detection and pose estimation from a single-view image is challenging due to the high uncertainty caused by the absence of 3D perception. As a solution, recent monocular 3D detection methods leverage additional modalities, such as stereo image pairs and LiDAR point clouds, to enhance image features at the expense of additional annotation costs. We propose using diffusion models to learn effective representations for monocular 3D detection without additional modalities or training data. We present $MonoDiff$, a novel framework that employs the reverse diffusion process to estimate 3D bounding box and orientation. But, considering the variability in bounding box sizes along different dimensions, it is ineffective to sample noise from a standard Gaussian distribution. Hence, we adopt a Gaussian mixture model to sample noise during the forward diffusion process and initialize the reverse diffusion process. Furthermore, since the diffusion model generates the 3D parameters for a given object image, we leverage 2D detection information to provide additional supervision by maintaining the correspondence between 3D/2D projection. Finally, depending on the signal-to-noise ratio, we incorporate a dynamic weighting scheme to account for the level of uncertainty in the supervision by projection at different timesteps. $MonoDiff$ outperforms current state-of-the-art monocular 3D detection methods on the KITTI and Waymo benchmarks without additional depth priors.

Reconstructing 3D clothed human involves creating a detailed geometry of individuals in clothing, with applications ranging from virtual try-on, movies, to games. To enable practical and widespread applications, recent advances propose to generate a clothed human from an RGB image. However, they struggle to reconstruct detailed and robust avatars simultaneously. We empirically find that the high-frequency (HF) and low-frequency (LF) information from a parametric model has the potential to enhance geometry details and improve robustness to noise, respectively. Based on this, we propose HiLo, namely clothed human reconstruction with high- and low-frequency information, which contains two components. 1) To recover detailed geometry using HF information, we propose a progressive HF Signed Distance Function to enhance the detailed 3D geometry of a clothed human. We analyze that our progressive learning manner alleviates large gradients that hinder model convergence. 2) To achieve robust reconstruction against inaccurate estimation of the parametric model by using LF information, we propose a spatial interaction implicit function. This function effectively exploits the complementary spatial information from a low-resolution voxel grid of the parametric model. Experimental results demonstrate that HiLo outperforms the state-of-the-art methods by 10.43% and 9.54% in terms of Chamfer distance on the Thuman2.0 and CAPE datasets, respectively. Additionally, HiLo demonstrates robustness to noise from the parametric model, challenging poses, and various clothing styles.

Recent advancements in single image driven 3D content generation have been propelled by leveraging prior knowledge from pretrained 2D diffusion models. However, the 3D content generated by existing methods often exhibits distorted outline shapes and inadequate details. To solve this problem, we propose a novel framework called Mask-enhanced Progressive Outline-to-Detail optimization (aka. MPOD123), which consists of two stages. Specifically, in the first stage, MPOD123 utilizes the pretrained view-conditioned diffusion model to guide the outline shape optimization of the 3D content. Given certain viewpoint, we estimate outline shape priors in the form of 2D mask from the 3D content by leveraging opacity calculation. In the second stage, MPOD123 incorporates Detail Appearance Inpainting (DAI) to guide the refinement on local geometry and texture with the shape priors. The essence of DAI lies in the Mask Rectified Cross-Attention (MRCA), which can be conveniently plugged in the stable diffusion model. The MRCA module utilizes the mask to rectify the attention map from each cross-attention layer. Accompanied with this new module, DAI is capable of guiding the detail refinement of the 3D content, while better preserves the outline shape. To assess the applicability in practical scenarios, we contribute a new dataset modeled on real-world e-commerce environments. Extensive quantitative and qualitative experiments on this dataset and open benchmarks demonstrate the effectiveness of MPOD123 over the state-of-the-arts.

Object pose refinement is essential for robust object pose estimation. Previous work has made significant progress towards instance-level object pose refinement. Yet, category-level pose refinement is a more challenging problem due to large shape variations within a category and the discrepancies between the target object and the shape prior. To address these challenges, we introduce a novel architecture for category-level object pose refinement. Our approach integrates an HS-layer and learnable affine transformations, which aims to enhance the extraction and alignment of geometric information. Additionally, we introduce a cross-cloud transformation mechanism that efficiently merges diverse data sources. Finally, we push the limits of our model by incorporating the shape prior information for translation and size error prediction. We conducted extensive experiments to demonstrate the effectiveness of the proposed framework. Through extensive quantitative experiments, we demonstrate significant improvement over the baseline method by a large margin across all metrics.

Understanding 3D object structure from image collections of general object categories remains a long-standing challenge in computer vision. Due to the high relevance of image keypoints (e.g. for graph matching, controlling generative models, scene understanding, etc.), in this work we specifically focus on inferring 3D structure in terms of sparse keypoints. Existing 3D keypoint inference approaches rely on strong priors, such as spatio-temporal consistency, multi-view images of the same object, 3D shape priors (e.g. templates, skeleton), or supervisory signals e.g. in the form of 2D keypoint annotations. In contrast, we propose the first unsupervised 3D keypoint inference approach that can be trained for general object categories solely from an inhomogeneous image collection (containing different instances of objects from the same category). Our experiments show that our method not only improves upon unsupervised 2D keypoint inference, but more importantly, it also produces reasonable 3D structure for various object categories, both qualitatively and quantitatively.

Reconstructing dynamic objects from monocular videos is a severely underconstrained and challenging problem, and recent work has approached it in various directions. However, owing to the ill-posed nature of this problem, there has been no solution that can provide consistent, high-quality novel views from camera positions that are significantly different from the training views. In this work, we introduce Neural Parametric Gaussians (NPGs) to take on this challenge by imposing a two-stage approach: first, we fit a low-rank neural deformation model, which then is used as regularization for non-rigid reconstruction in the second stage. The first stage learns the object’s deformations such that it preserves consistency in novel views. The second stage obtains high reconstruction quality by optimizing 3D Gaussians that are driven by the coarse model. To this end, we introduce a local 3D Gaussian representation, where temporally shared Gaussians are anchored in and deformed by local oriented volumes. The resulting combined model can be rendered as radiance fields, resulting in high-quality photo-realistic reconstructions of the non-rigidly deforming objects, maintaining 3D consistency across novel views. We demonstrate that NPGs achieve superior results compared to previous works, especially in challenging scenarios with few multi-view cues.

Creating personalized hand avatars is important to offer a realistic experience to users on AR / VR platforms. While most prior studies focused on reconstructing 3D hand shapes, some recent work has tackled the reconstruction of hand textures on top of shapes. However, these methods are often limited to capturing pixels on the visible side of a hand, requiring diverse views of the hand in a video or multiple images as input. In this paper, we propose a novel method, BiTT(Bi-directional Texture reconstruction of Two hands), which is the first end-to-end trainable method for relightable, pose-free texture reconstruction of two interacting hands taking only a single RGB image, by three novel components: 1)\ bi-directional (left $\leftrightarrow$ right) texture reconstruction using the texture symmetry of left / right hands, 2) utilizing a texture parametric model for hand texture recovery, and 3)\ the overall coarse-to-fine stage pipeline for reconstructing personalized texture of two interacting hands. BiTT first estimates the scene light condition and albedo image from an input image, then reconstructs the texture of both hands through the texture parametric model and bi-directional texture reconstructor. In experiments using InterHand2.6M and RGB2Hands datasets, our method significantly outperforms state-of-the-art hand texture reconstruction methods quantitatively and qualitatively. The code and data will be publicly available.

In this paper, we introduce a novel approach that harnesses both 2D and 3D attentions to enable highly accurate depth completion without requiring iterative spatial propagations. Specifically, we first enhance a baseline convolutional depth completion model by applying attention to 2D features in the bottleneck and skip connections. This effectively improves the performance of this simple network and sets it on par with the latest, complex transformer-based models. Leveraging the initial depths and features from this network, we uplift the 2D features to form a 3D point cloud and construct a 3D point transformer to process it, allowing the model to explicitly learn and exploit 3D geometric features. In addition, we propose normalization techniques to process the point cloud, which improves learning and leads to better accuracy than directly using point transformers off the shelf. Furthermore, we incorporate global attention on downsampled point cloud features, which enables long-range context while still being computationally feasible. We evaluate our method, DeCoTR, on established depth completion benchmarks, including NYU Depth V2 and KITTI, showcasing that it sets new state-of-the-art performance. We further conduct zero-shot evaluations on ScanNet and DDAD benchmarks and demonstrate that DeCoTR has superior generalizability compared to existing approaches.

We present Monocular Neural Parametric Head Models (MonoNPHM) for dynamic 3D head reconstructions from monocular RGB videos. To this end, we propose a latent appearance space that parameterizes a texture field on top of a neural parametric model. We constrain predicted color values to be correlated with the underlying geometry such that gradients from RGB effectively influence latent geometry codes during inverse rendering. To increase the representational capacity of our expression space, we augment our backward deformation field with hyper-dimensions, thus improving color and geometry representation in topologically challenging expressions. Using MonoNPHM as a learned prior, we approach the task of 3D head reconstruction using signed distance field based volumetric rendering. By numerically inverting our backward deformation field, we incorporated a landmark loss using facial anchor points that are closely tied to our canonical geometry representation. To evaluate the task of dynamic face reconstruction from monocular RGB videos we record 20 challenging Kinect sequences under casual conditions. MonoNPHM outperforms all baselines with a significant margin, and makes an important step towards easily accessible neural parametric face models through RGB tracking.

Due to the high potential for abuse, the task of detecting synthetic images has lately become of great interest to the research community. Unfortunately, existing image-space detectors quickly become obsolete as new high-fidelity text-to-image models are developed at blinding speed. In this work, we propose a new synthetic image detector that uses features obtained by inverting an open-source pre-trained Stable Diffusion model. We show that these inversion features enable our detector to generalize well to unseen generators of high visual fidelity (e.g., DALL·E 3) even when the detector is trained only on lower fidelity fake images generated via Stable Diffusion.We show that the resulting detector achieves new state-of-the-art across multiple training and evaluation setups. Moreover, we introduce a new challenging evaluation protocol that uses reverse image search to remove stylistic and thematic biases from the detector evaluation. We show that the resulting evaluation scores aligns well with detectors' in-the-wild performance, and release these datasets as pubic benchmarks for future research.

In this paper, we study the problem of generalizable synthetic image detection, aiming to detect forgery images from diverse generative methods, e.g., GANs and diffusion models. Cutting-edge solutions start to explore the benefits of pre-trained models, and mainly follow the fixed paradigm of solely training an attached classifier, e.g., combining frozen CLIP-ViT with a learnable linear layer in UniFD [35]. However, our analysis shows that such a fixed paradigm is prone to yield detectors with insufficient learning regarding forgery representations. We attribute the key challenge to the lack of forgery adaptation, and present a novel forgery-aware adaptive transformer approach, namely FatFormer. Based on the pre-trained vision-language spaces of CLIP, FatFormer introduces two core designs for the adaption to build generalized forgery representations. First, motivated by the fact that both image and frequency analysis are essential for synthetic image detection, we develop a forgery-aware adapter to adapt image features to discern and integrate local forgery traces within image and frequency domains. Second, we find that considering the contrastive objectives between adapted image features and text prompt embeddings, a previously overlooked aspect, results in a nontrivial generalization improvement. Accordingly, we introduce language-guided alignment to supervise the forgery adaptation with image and text prompts in FatFormer. Experiments show that, by coupling these two designs, our approach tuned on 4-class ProGAN data attains a remarkable detection performance, achieving an average of 98% accuracy to unseen GANs, and surprisingly generalizes to unseen diffusion models with 95% accuracy.

In recent years, image manipulation localization has attracted increasing attention due to its pivotal role in guaranteeing social media security. However, how to accurately identify the forged regions remains an open challenge. One of the main bottlenecks lies in the severe scarcity of high-quality data, due to its costly creation process. To address this limitation, we propose a novel paradigm, termed as CAAA, to automatically and precisely annotate the numerous manually forged images from the web at the pixel level. We further propose a novel metric QES to facilitate the automatic filtering of unreliable annotations. With CAAA and QES, we construct a large-scale, diverse, and high-quality dataset comprising 123,150 manually forged images with mask annotations. Besides, we develop a new model APSC-Net for accurate image manipulation localization. According to extensive experiments, our dataset significantly improves the performance of various models on the widely-used benchmarks and such improvements are attributed to our proposed effective methods. The dataset and code are publicly available at https://github.com/qcf-568/MIML.

Recent works have shown that generative models leave traces of their underlying generative process on the generated samples, broadly referred to as fingerprints of a generative model, and have studied their utility in detecting synthetic images from real ones. However, the extend to which these fingerprints can distinguish between various types of synthetic image and help identify the underlying generative process remain under-explored. In particular, the very definition of a fingerprint remains unclear, to our knowledge. To that end, in this work, we formalize the definition of artifact and fingerprint in generative models, propose an algorithm for computing them in practice, and finally study its effectiveness in distinguishing a large array of different generative models. We find that using our proposed definition can significantly improve the performance on the task of identifying the underlying generative process from samples (model attribution) compared to existing methods. Additionally, we study the structure of the fingerprints, and observe that it is very predictive of the effect of different design choices on the generative process.

Generative AI (GenAI) is transforming creative workflows through the capability to synthesize and manipulate images via high-level prompts. Yet creatives are not well supported to receive recognition or reward for the use of their content in GenAI training. To this end, we propose ProMark, a causal attribution technique to attribute a synthetically generated image to its training data concepts like objects, motifs, templates, artists, or styles. The concept information is proactively embedded into the input training images using imperceptible watermarks, and the diffusion models (unconditional or conditional) are trained to retain the corresponding watermarks in generated images. We show that we can embed as many as $2^{16}$ unique watermarks into the training data, and each training image can contain more than one watermark. ProMark can maintain image quality whilst outperforming correlation-based attribution. Finally, several qualitative examples are presented, providing the confidence that the presence of the watermark conveys a causative relationship between training data and synthetic images.

Diffusion Models (DMs) have evolved into advanced image generation tools, especially for few-shot generation where a pre-trained model is fine-tuned on a small set of images to capture a specific style or object. Despite their success, concerns exist about potential copyright violations stemming from the use of unauthorized data in this process. In response, we present Contrasting Gradient Inversion for Diffusion Models (CGI-DM), a novel method featuring vivid visual representations for digital copyright authentication. Our approach involves removing partial information of an image and recovering missing details by exploiting conceptual differences between the pre-trained and fine-tuned models. We formulate the differences as KL divergence between latent variables of the two models when given the same input image, which can be maximized through Monte Carlo sampling and Projected Gradient Descent (PGD). The similarity between original and recovered images serves as a strong indicator of potential infringements. Extensive experiments on the WikiArt and Dreambooth datasets demonstrate the high accuracy of CGI-DM in digital copyright authentication, surpassing alternative validation techniques. Code implementation is available at https://github.com/Nicholas0228/Revelio.

Accurate representation in media is known to improve the well-being of the people who consume it. Generative image models trained on large web-crawled datasets such as LAION are known to produce images with harmful stereotypes and misrepresentations of cultures. We improve inclusive representation in generated images by (1) engaging with communities to collect a culturally representative dataset that we call the Cross-Cultural Understanding Benchmark (CCUB), and we propose (2) a novel Self-Contrastive Fine-Tuning (SCoFT) method that leverages the model's known biases to self-improve. SCoFT is designed to encode high-level information from the dataset into the model for the purpose of shifting away from misrepresentations of a culture.Our user study conducted on 51 participants from 5 different countries based on their self-selected national cultural affiliation shows that our proposed approach consistently generates images with higher cultural relevance and fewer stereotypes when compared to the Stable Diffusion baseline.

This paper investigates the impact of recent deep generative models, such as stable diffusion, on potential biases in upcoming models. As the internet witnesses an increasing influx of images generated by these models, concerns arise regarding inherent biases that may accompany them, potentially leading to the dissemination of harmful content. The primary question explored in this study is whether a detrimental feedback loop, resulting in the amplification of bias, would ensue if these generated images were used as the training data for future models. Toward this concern, we conduct simulations by progressively substituting original images in COCO and CC3M datasets with images generated through stable diffusion. Subsequently, these modified datasets are utilized to train image caption and CLIP models, and the resulting models are evaluated using both quality and bias metrics. Contrary to expectations, our findings indicate that the introduction of generated images during training does not uniformly amplify bias. Instead, instances of bias mitigation across specific tasks are observed. Our exploration extends to identifying factors influencing bias changes, such as features in image generation (e.g., producing blurry faces) and pre-existing biases in the original dataset.

Diffusion models have demonstrated unprecedented capabilities in image generation. Yet, they incorporate and amplify the data bias (e.g., gender, age) from the original training set, limiting the diversity of generated images. In this paper, we propose a diversity-oriented fine-tuning method using reinforcement learning (RL) for diffusion models under the guidance of an image-set-based reward function. Specifically, the proposed reward function, denoted as Diversity Reward, utilizes a set of generated images to evaluate the coverage of the current generative distribution w.r.t. the reference distribution, represented by a set of unbiased images. Built on top of the probabilistic method of distribution discrepancy estimation, Diversity Reward can measure the relative distribution gap with a small set of images efficiently. We further formulate the diffusion process as a multi-step decision-making problem (MDP) and apply policy gradient methods to fine-tune diffusion models by maximizing the Diversity Reward. The proposed rewards are validated on a post-sampling selection task, where a subset of the most diverse images are selected based on Diversity Reward values. We also show the effectiveness of our RL fine-tuning framework on enhancing the diversity of image generation with different types of diffusion models, including class-conditional models and text-conditional models, e.g., StableDiffusion.

The goal of selective prediction is to allow an a model to abstain when it may not be able to deliver a reliable prediction, which is important in safety-critical contexts. Existing approaches to selective prediction typically require access to the internals of a model, require retraining a model or study only unimodal models. However, the most powerful models (e.g. GPT-4) are typically only available as black boxes with inaccessible internals, are not retrainable by end-users, and are frequently used for multimodal tasks. We study the possibility of selective prediction for vision-language models in a realistic, black-box setting. We propose using the principle of neighborhood consistency to identify unreliable responses from a black-box vision-language model in question answering tasks. We hypothesize that given only a visual question and model response, the consistency of the model's responses over the neighborhood of a visual question will indicate reliability. It is impossible to directly sample neighbors in feature space in a black-box setting. Instead, we show that it is possible to use a smaller proxy model to approximately sample from the neighborhood. We find that neighborhood consistency can be used to identify model responses to visual questions that are likely unreliable, even in adversarial settings or settings that are out-of-distribution to the proxy model.

In film gender studies, the concept of “male gaze” refers to the way the characters are portrayed on-screen as objects of desire rather than subjects. In this article, we introduce a novel video-interpretation task, to detect character objectification in films. The purpose is to reveal and quantify the usage of complex temporal patterns operated in cinema to produce the cognitive perception of objectification.We introduce the ObyGaze12 dataset, made of 1914 movie clips densely annotated by experts for objectification concepts identified in film studies and psychology.We evaluate recent vision models, show the feasibility of the task and where the challenges remain with concept bottleneck models. Our new dataset and code are made available to the community.

The rapid evolution of automatic facial indexing technologies increases the risk of compromising personal and sensitive information. To mitigate the issue, we propose creating cartoon avatars, or 'toon avatars', designed to effectively obscure identity features. The primary objective is to deceive current AI systems, preventing them from accurately identifying individuals while making minimal modifications to their facial features. Moreover, we aim to ensure that a human observer can still recognize the person depicted in these altered avatar images. To achieve this, we introduce 'ToonerGAN', a novel approach that utilizes Generative Adversarial Networks (GANs) to craft personalized cartoon avatars. The ToonerGAN framework consists of a style module and a de-identification module that work together to produce high-resolution, realistic cartoon images. For the efficient training of our network, we have developed ‘ToonSet' dataset, consisting of around 23,000 facial images and their cartoon renditions. Through comprehensive experiments and benchmarking against existing datasets, including CelebA-HQ, our method demonstrates superior performance in obfuscating identity while preserving the utility of data. Additionally, a user-centric study exploring the effectiveness of ToonerGAN has yielded compelling observations.

Recent advancements in post-hoc and inherently interpretable methods have markedly enhanced the explanations of black box classifier models. These methods operate either through post-analysis or by integrating concept learning during model training. Although being effective in bridging the semantic gap between a model's latent space and human interpretation, these explanation methods only partially reveal the model's decision-making process. The outcome is typically limited to high-level semantics derived from the last feature map. We argue that the explanations lacking insights into the decision processes at low and mid-level features are neither fully faithful nor useful. Addressing this gap, we introduce the Multi-Level Concept Prototypes Classifier (MCPNet), an inherently interpretable model. MCPNet autonomously learns meaningful concept prototypes across multiple feature map levels using Centered Kernel Alignment (CKA) loss and an energy-based weighted PCA mechanism, and it does so without reliance on predefined concept labels. Further, we propose a novel classifier paradigm that learns and aligns multi-level concept prototype distributions for classification purposes. Our experiments reveal that our proposed MCPNet, while being adaptable to various model architectures, offers comprehensive multi-level explanations with maintaining the classification accuracy. Additionally, its concept distribution-based classification approach shows improved generalization capabilities in few-shot classification scenarios.

Understanding what deep network models capture in their learned representations is a fundamental challenge in computer vision. We present a new methodology to understanding such vision models, the Visual Concept Connectome (VCC), which discovers human interpretable concepts and their interlayer connections in a fully unsupervised manner. Our approach simultaneously reveals fine-grained concepts at a layer, connection weightings across all layers and is amendable to global analysis of network structure (e.g., branching pattern of hierarchical concept assemblies). Previous work yielded ways to extract interpretable concepts from single layers and examine their impact on classification, but did not afford multilayer concept analysis across an entire network architecture. Quantitative and qualitative empirical results show the effectiveness of VCCs in the domain of image classification.

Machine learning models can perform well on in-distribution data but often fail on biased subgroups that are underrepresented in the training data, hindering the robustness of models for reliable applications. Such subgroups are typically unknown due to the absence of subgroup labels. Discovering biased subgroups is the key to understanding models' failure modes and further improving models' robustness. Most previous works of subgroup discovery make an implicit assumption that models only underperform on a single biased subgroup, which does not hold on in-the-wild data where multiple biased subgroups exist. In this work, we propose Decomposition, Interpretation, and Mitigation (DIM), a novel method to address a more challenging but also more practical problem of discovering multiple biased subgroups in image classifiers. Our approach decomposes the image features into multiple components that represent multiple subgroups. This decomposition is achieved via a bilinear dimension reduction method, Partial Least Square (PLS), guided by useful supervision from the image classifier. We further interpret the semantic meaning of each subgroup component by generating natural language descriptions using vision-language foundation models. Finally, DIM mitigates multiple biased subgroups simultaneously via two strategies, including the data- and model-centric strategies. Extensive experiments on CIFAR-100 and Breeds datasets demonstrate the effectiveness of DIM in discovering and mitigating multiple biased subgroups. Furthermore, DIM uncovers the failure modes of the classifier on Hard ImageNet, showcasing its broader applicability to understanding model bias in image classifiers.

Deep neural networks (DNNs) often display overconfidence when encountering out-of-distribution (OOD) samples, posing significant challenges in real-world applications. Capitalizing on the observation that responses on convolutional kernels are generally more pronounced for in-distribution (ID) samples than for OOD ones, this paper proposes the COnvolutional REsponse-based Score (CORES) to exploit these discrepancies for OOD detection. Initially, CORES delves into the extremities of convolutional responses by considering both their magnitude and the frequency of significant values. Moreover, through backtracking from the most prominent predictions, CORES effectively pinpoints sample-relevant kernels across different layers. These kernels, which exhibit a strong correlation to input samples, are integral to CORES's OOD detection capability. Comprehensive experiments across various ID and OOD settings demonstrate CORES's effectiveness in OOD detection and its superiority to the state-of-the-art methods.

While Transformers have rapidly gained popularity in various computer vision applications, post-hoc explanations of their internal mechanisms remain largely unexplored. Vision Transformers extract visual information by representing image regions as transformed tokens and integrating them via attention weights. However, existing post-hoc explanation methods merely consider these attention weights, neglecting crucial information from the transformed tokens, which fails to accurately illustrate the rationales behind the models' predictions. To incorporate the influence of token transformation into interpretation, we propose TokenTM, a novel post-hoc explanation method that utilizes our introduced measurement of token transformation effects. Specifically, we quantify token transformation effects by measuring changes in token lengths and correlations in their directions pre- and post-transformation. Moreover, we develop initialization and aggregation rules to integrate both attention weights and token transformation effects across all layers, capturing holistic token contributions throughout the model. Experimental results on segmentation and perturbation tests demonstrate the superiority of our proposed TokenTM compared to state-of-the-art Vision Transformer explanation methods.

In order to gain insights about the decision-making of different visual recognition backbones, we propose two methodologies, sub-explanation counting and cross-testing, that systematically applies deep explanation algorithms on a dataset-wide basis, and compares the statistics generated from the amount and nature of the explanations. These methodologies reveal the difference among networks in terms of two properties called compositionality and disjunctivism. Transformers and ConvNeXt are found to be more compositional, in the sense that they jointly consider multiple parts of the image in building their decisions, whereas traditional CNNs and distilled transformers are less compositional and more disjunctive, which means that they use multiple diverse but smaller set of parts to achieve a confident prediction. Through further experiments, we pinpointed the choice of normalization to be especially important in the compositionality of a model, in that batch normalization leads to less compositionality while group and layer normalization lead to more. Finally, we also analyze the features shared by different backbones and plot a landscape of different models based on their feature-use similarity.

To interpret Vision Transformers, post-hoc explanations assign salience scores to input pixels, providing human-understandable heatmaps. However, whether these interpretations reflect true rationales behind the model's output is still underexplored.To address this gap, we study the faithfulness criterion of explanations: the assigned salience scores should represent the influence of the corresponding input pixels on the model's predictions. To evaluate faithfulness, we introduce Salience-guided Faithfulness Coefficient (SaCo), a novel evaluation metric leveraging essential information of salience distribution. Specifically, we conduct pair-wise comparisons among distinct pixel groups and then aggregate the differences in their salience scores, resulting in a coefficient that indicates the explanation's degree of faithfulness. Our explorations reveal that current metrics struggle to differentiate between advanced explanation methods and Random Attribution, thereby failing to capture the faithfulness property. In contrast, our proposed SaCo offers a reliable faithfulness measurement, establishing a robust metric for interpretations. Furthermore, our SaCo demonstrates that the use of gradient and multi-layer aggregation can markedly enhance the faithfulness of attention-based explanation, shedding light on potential paths for advancing Vision Transformer explainability.

This paper studies the problem of concept-based interpretability of transformer representations for videos. Concretely, we seek to explain the decision-making process of video transformers based on high-level, spatiotemporal concepts that are automatically discovered. Prior research on concept-based interpretability has concentrated solely on image-level tasks, like image classification. Comparatively, video models deal with the added temporal dimension, increasing complexity and posing challenges in identifying dynamic concepts over time.In this work, we systematically address these challenges by introducing the first Video Transformer Concept Discovery (VTCD) algorithm. To this end, we propose an efficient approach for unsupervised identification of units of video transformer representations - concepts. We then design a noise-robust algorithm for ranking the importance of these units to the output of a model, allowing us to analyze its decision making process. Performing this analysis jointly over a diverse set of supervised and self-supervised models we make a number of important discoveries about universal units of video representations. Finally, we demonstrate that VTCD can be used to improve model performance for fine-grained tasks.

The many variations of Implicit Neural Representations (INRs), where a neural network is trained as a continuous representation of a signal, have tremendous practical utility for downstream tasks including novel view synthesis, video compression, and image superresolution. Unfortunately, the inner workings of these networks are seriously under-studied. Our work, eXplaining the Implicit Neural Canvas (XINC), is a unified framework for explaining properties of INRs by examining the strength of each neuron’s contribution to each output pixel. We call the aggregate of these contribution maps the Implicit Neural Canvas and we use this concept to demonstrate that the INRs which we study learn to “see” the frames they represent in surprising ways. For example, INRs tend to have highly distributed representations. While lacking high-level object semantics, they have a significant bias for color and edges, and are almost entirely space-agnostic. We arrive at our conclusions by examining how objects are represented across time in video INRs, using clustering to visualize similar neurons across layers and architectures, and show that this is dominated by motion. These insights demonstrate the general usefulness of our analysis framework.

Recent advancements in neural networks have showcased their remarkable capabilities across various domains. Despite these successes, the “black box” problemstill remains. Addressing this, we propose a novel framework, WWW, that offers the ‘what’, ‘where’, and ‘why’of the neural network decisions in human-understandableterms. Specifically, WWW utilizes adaptive selection forconcept discovery, employing adaptive cosine similarityand thresholding techniques to effectively explain ‘what’.To address the ‘where’ and ‘why’, we proposed a novelcombination of neuron activation maps (NAMs) with Shapleyvalues, generating localized concept maps and heatmapsfor individual inputs. Furthermore, WWW introduces amethod for predicting uncertainty, leveraging heatmap similarities to estimate ‘how’ reliable the prediction is. Experimental evaluations of WWW demonstrate superior performance in both quantitative and qualitative metrics, outperforming existing methods in interpretability. WWW provides a unified solution for explaining ‘what’, ‘where’, and‘why’, introducing a method for localized explanations fromglobal interpretations and offering a plug-and-play solutionadaptable to various architectures.

This paper addresses the decomposition of holographic feature vectors in Hyperdimensional Computing (HDC) aka Vector Symbolic Architectures (VSA). HDC uses high-dimensional vectors with brain-like properties to represent symbolic information, and leverages efficient operators to construct and manipulate complexly structured data in a cognitive fashion. Existing models face challenges in decomposing these structures, a process crucial for understanding and interpreting a composite hypervector. We address this challenge by proposing the HDC Memorized-Factorization Problem that captures the common patterns of construction in HDC models. To solve this problem efficiently, we introduce HDQMF, a HyperDimensional Quantum Memorized-Factorization algorithm. HDQMF is unique in its approach, utilizing quantum computing to offer efficient solutions. It modifies crucial steps in Grover's algorithm to achieve hypervector decomposition, achieving quadratic speed-up.

Local Interpretable Model-agnostic Explanations (LIME) - a widely used post-ad-hoc model agnostic explainable AI (XAI) technique. It works by training a simple transparent (surrogate) model using random samples drawn around the neighborhood of the instance (image) to be explained (IE). Explanations are then extracted for a black-box model and a given IE, using the surrogate model. However, the explanations of LIME suffer from inconsistency across different runs for the same model and the same IE. We identify two main types of inconsistencies: variance in the sign and importance ranks of the segments (superpixels). These factors hinder LIME from obtaining consistent explanations. We analyze these inconsistencies and propose a new method, Stabilized LIME for Consistent Explanations (SLICE). The proposed method handles the stabilization problem in two aspects: using a novel feature selection technique to eliminate spurious superpixels and an adaptive perturbation technique to generate perturbed images in the neighborhood of IE. Our results demonstrate that the explanations from SLICE exhibit significantly better consistency and fidelity than LIME (and its variant BayLime).

In this paper, we explore the unique modality of sketch for explainability, emphasising the profound impact of human strokes compared to conventional pixel-oriented studies. Beyond explanations of network behavior, we discern the genuine implications of explainability across diverse downstream sketch-related tasks. We propose a lightweight and portable explainability solution -- a seamless plugin that integrates effortlessly with any pre-trained model, eliminating the need for re-training. Demonstrating its adaptability, we present four applications: highly studied retrieval and generation, and completely novel assisted drawing and sketch adversarial attacks. The centrepiece to our solution is a stroke-level attribution map that takes different forms when linked with downstream tasks. By addressing the inherent non-differentiability of rasterisation, we enable explanations at both coarse stroke level (SLA) and partial stroke level (P-SLA), each with its advantages for specific downstream tasks.

Gradient-based saliency maps have been widely used to explain the decisions of deep neural network classifiers. However, standard gradient-based interpretation maps, including the simple gradient and integrated gradient algorithms, often lack desired structures such as sparsity and connectedness in their application to real-world computer vision models. A common approach to induce sparsity-based structures into gradient-based saliency maps is to modify the simple gradient scheme using sparsification or norm-based regularization. However, one drawback with such post-processing approaches is the potentially significant loss in fidelity to the original simple gradient map. In this work, we propose to apply adversarial training as an in-processing scheme to train neural networks with structured simple gradient maps. We demonstrate an existing duality between the regularized norms of the adversarial perturbations and gradient-based maps, whereby we design adversarial training schemes promoting sparsity and group-sparsity properties in simple gradient maps. We present comprehensive numerical results to show the influence of our proposed norm-based adversarial training methods on the standard gradient-based maps of standard neural network architectures on benchmark image datasets.

Convolution neural network is successful in pervasive vision tasks, including label distribution learning, which usually takes the form of learning an injection from the nonlinear visual features to the well-defined labels. However, how the discrepancy between features is mapped to the label discrepancy is ambient, and its correctness is not guaranteed. To address these problems, we study the mathematicalconnection between feature and its label, presenting a general and simple framework for label distribution learning. We propose a so-called Triangular Distribution Transform (TDT) to build an injective function between feature and label, guaranteeing that any symmetric feature discrepancy linearly reflects the difference between labels. The proposed TDT can be used as a plug-in in mainstream backbone networks to address different label distribution learning tasks. Experiments on Facial Age Recognition, Illumination Chromaticity Estimation, and Aesthetics assessment show that TDT achieves on-par or better results than the prior arts. Code is available at https://github.com/redcping/TDT.

Concept Bottleneck Models (CBMs) map the black-box visual representations extracted by deep neural networks onto a set of interpretable concepts and use the concepts to make predictions, enhancing the transparency of the decision-making process. Multimodal pre-trained models can match visual representations with textual concept embeddings, allowing for obtaining the interpretable concept bottleneck without the expertise concept annotations. Recent research has focused on the concept bank establishment and the high-quality concept selection. However, it is challenging to construct a comprehensive concept bank through humans or large language models, which severely limits the performance of CBMs. In this work, we propose the Incremental Residual Concept Bottleneck Model (Res-CBM) to address the challenge of concept completeness. Specifically, the residual concept bottleneck model employs a set of optimizable vectors to complete missing concepts, then the incremental concept discovery module converts the complemented vectors with unclear meanings into potential concepts in the candidate concept bank. Our approach can be applied to any user-defined concept bank, as a post-hoc processing method to enhance the performance of any CBMs. Furthermore, to measure the descriptive efficiency of CBMs, the Concept Utilization Efficiency (CUE) metric is proposed. Experiments show that the Res-CBM outperforms the current state-of-the-art methods in terms of both accuracy and efficiency and achieves comparable performance to black-box models across multiple datasets.

In ill-posed inverse problems, it is commonly desirable to obtain insight into the full spectrum of plausible solutions, rather than extracting only a single reconstruction. Information about the plausible solutions and their likelihoods is encoded in the posterior distribution. However, for high-dimensional data, this distribution is challenging to visualize. In this work, we introduce a new approach for estimating and visualizing posteriors by employing energy-based models (EBMs) over low-dimensional subspaces. Specifically, we train a conditional EBM that receives an input measurement and a set of directions that span some low-dimensional subspace of solutions, and outputs the probability density function of the posterior within that space. We demonstrate the effectiveness of our method across a diverse range of datasets and image restoration problems, showcasing its strength in uncertainty quantification and visualization. As we show, our method outperforms a baseline that projects samples from a diffusion-based posterior sampler, while being orders of magnitude faster. Furthermore, it is more accurate than a baseline that assumes a Gaussian posterior.

Epistemic uncertainty quantification (UQ) identifies where models lack knowledge. Traditional UQ methods, often based on Bayesian neural networks, are not suitable for pre-trained non-Bayesian models. Our study addresses quantifying epistemic uncertainty for any pre-trained model, which does not need the original training data or model modifications and can ensure broad applicability regardless of network architectures or training techniques. Specifically, we propose a gradient-based approach to assess epistemic uncertainty, analyzing the gradients of outputs relative to model parameters, and thereby indicating necessary model adjustments to accurately represent the inputs. We first explore theoretical guarantees of gradient-based methods for epistemic UQ, questioning the view that this uncertainty is only calculable through differences between multiple models. We further improve gradient-driven UQ by using class-specific weights for integrating gradients and emphasizing distinct contributions from neural network layers. Additionally, we enhance UQ accuracy by combining gradient and perturbation methods to refine the gradients. We evaluate our approach on out-of-distribution detection, uncertainty calibration, and active learning, demonstrating its superiority over current state-of-the-art UQ methods for pre-trained models.

Quantifying the degree of similarity between images is a key copyright issue for image-based machine learning. In legal doctrine however, determining the degree of similarity between works requires subjective analysis, and fact-finders (judges and juries) can demonstrate considerable variability in these subjective judgement calls. Images that are structurally similar can be deemed dissimilar, whereas images of completely different scenes can be deemed similar enough to support a claim of copying. We seek to define and compute a notion of "conceptual similarity" among images that captures high-level relations even among images that do not share repeated elements or visually similar components. The idea is to use a base multi-modal model to generate "explanations" (captions) of visual data at increasing levels of complexity. Then, similarity can be measured by the length of the caption needed to discriminate between the two images: Two highly dissimilar images can be discriminated early in their description, whereas conceptually dissimilar ones will need more detail to be distinguished. We operationalize this definition and show that it correlates with subjective (averaged human evaluation) assessment, and beats existing baselines on both image-to-image and text-to-text similarity benchmarks. Beyond just providing a number, our method also offers interpretability by pointing to the specific level of granularity of the description where the source data is differentiated.

Deep Neural Networks (DNNs) are widely used for visual classification tasks, but their complex computation process and black-box nature hinder decision transparency and interpretability. Class activation maps (CAMs) and recent variants provide ways to visually explain the DNN decision-making process by displaying `attention' heatmaps of the DNNs. Nevertheless, the CAM explanation only offers relative attention information, that is, on an attention heatmap, we can interpret which image region is more or less important than the others. However, these regions cannot be meaningfully compared across classes, and the contribution of each region to the model's class prediction is not revealed. To address these challenges that ultimately lead to better DNN Interpretation, in this paper, we propose CAPE, a novel reformulation of CAM that provides a unified and probabilistically meaningful assessment of the contributions of image regions. We quantitatively and qualitatively compare CAPE with state-of-the-art CAM methods on CUB and ImageNet benchmark datasets to demonstrate enhanced interpretability. We also test on a cytology imaging dataset depicting a challenging Chronic Myelomonocytic Leukemia (CMML) diagnosis problem.

Addressing biases in computer vision models is crucial for real-world AI system deployments. However, mitigating visual biases is challenging due to their unexplainable nature, often identified indirectly through visualization or sample statistics, which necessitates additional human supervision for interpretation. To tackle this issue, we propose the Bias-to-Text (B2T) framework, which interprets visual biases as keywords. Specifically, we extract common keywords from the captions of mispredicted images to identify potential biases in the model. We then validate these keywords by measuring their similarity to the mispredicted images using a vision-language scoring model. The keyword explanation form of visual bias offers several advantages, such as a clear group naming for bias discovery and a natural extension for debiasing using these group names. Our experiments demonstrate that B2T can identify known biases, such as gender bias in CelebA, background bias in Waterbirds, and distribution shifts in ImageNet-R and ImageNet-C. Additionally, B2T uncovers novel biases in larger datasets, such as Dollar Street and ImageNet. For example, we discovered a contextual bias between "bee" and "flower" in ImageNet. We also highlight various applications of B2T keywords, including debiased training, CLIP prompting, model comparison, and label diagnosis.

While deep learning has led to huge progress in complex image classification tasks like ImageNet, unexpected failure modes, e.g. via spurious features, call into question how reliably these classifiers work in the wild. Furthermore, for safety-critical tasks the black-box nature of their decisions is problematic, and explanations or at least methods which make decisions plausible are needed urgently. In this paper, we address these problems by generating images that optimize a classifier-derived objective using a framework for guided image generation. We analyze the behavior and decisions of image classifiers by visual counterfactual explanations (VCEs), detection of systematic mistakes by analyzing images where classifiers maximally disagree, and visualization of neurons to verify potential spurious features. In this way, we validate existing observations, e.g. the shape bias of adversarially robust models, as well as novel failure modes, e.g. systematic errors of zero-shot CLIP classifiers, or identify harmful spurious features. Moreover, our VCEs outperform previous work while being more versatile.

Accurate 3D neuron segmentation from electron microscopy (EM) volumes is crucial for neuroscience research. However, the complex neuron morphology often leads to over-merge and over-segmentation results. Recent advancements utilize 3D CNNs to predict a 3D affinity map with improved accuracy but suffer from two challenges: high computational cost and limited input size, especially for practical deployment for large-scale EM volumes. To address these challenges, we propose a novel method to leverage lightweight 2D CNNs for efficient neuron segmentation. Our method employs a 2D Y-shape network to generate two embedding maps from adjacent 2D sections, which are then converted into an affinity map by measuring their embedding distance. While the 2D network better captures pixel dependencies inside sections with a larger input size, it overlooks inter-section dependencies. To overcome this, we introduce a cross-dimension affinity distillation (CAD) strategy that transfers inter-section dependency knowledge from a 3D teacher network to the 2D student network by ensuring consistency between their output affinity maps. Additionally, we design a feature grafting interaction (FGI) module to enhance knowledge transfer by grafting embedding maps from the 2D student onto those from the 3D teacher. Extensive experiments on multiple EM neuron segmentation datasets, including a newly built one by ourselves, demonstrate that our method achieves superior performance over state-of-the-art methods with only 1/20 inference latency.

Self-supervised learning (SSL) is an efficient pre-training method for medical image analysis. However, current research is mostly confined to certain modalities, consuming considerable time and resources without achieving universality across different modalities. A straightforward solution is combining all modality data for joint SSL, which poses practical challenges. Firstly, our experiments reveal conflicts in representation learning as the number of modalities increases. Secondly, multi-modal data collected in advance cannot cover all real-world scenarios. In this paper, we reconsider versatile SSL from the perspective of continual learning and propose MedCoSS, a continuous SSL approach for multi-modal medical data. Different from joint representation learning, MedCoSS assigns varying data modalities to separate training stages, creating a multi-stage pre-training process. We propose a rehearsal-based continual learning approach to manage modal conflicts and prevent catastrophic forgetting. Specifically, we use the k-means sampling to retain and rehearse previous modality data during new modality learning. Moreover, we apply feature distillation and intra-modal mixup on buffer data for knowledge retention, bypassing pretext tasks. We conduct experiments on a large-scale multi-modal unlabeled dataset, including clinical reports, X-rays, CT, MRI, and pathological images. Experimental results demonstrate MedCoSS’s exceptional generalization ability across 9 downstream datasets and its significant scalability in integrating new modality data. The code and pre-trained model are available at https://github.com/yeerwen/MedCoSS.

Defocus blur is a persistent problem in microscope imaging that poses harm to pathology interpretation and medical intervention in cell microscopy and microscope surgery. To address this problem, a unified framework including the multi-pyramid transformer (MPT) and extended frequency contrastive regularization (EFCR) is proposed to tackle two outstanding challenges in microscopy deblur: longer attention span and data deficiency. The MPT employs an explicit pyramid structure at each network stage that integrates the cross-scale window attention (CSWA), the intra-scale channel attention (ISCA), and the feature-enhancing feed-forward network (FEFN) to capture long-range cross-scale spatial interaction and global channel context. The EFCR addresses the data deficiency problem by exploring latent deblur signals from different frequency bands. It also enables deblur knowledge transfer to learn cross-domain information from extra data, improving deblur performance for labeled and unlabeled data. Extensive experiments and downstream task validation show the framework achieves state-of-the-art performance across multiple datasets. Project page: https://github.com/PieceZhang/MPT-CataBlur.

The advancement of Zero-Shot Learning in the medical domain has been driven forward by using pre-trained models on large-scale image-text pairs, focusing on image-text alignment. However, existing methods primarily rely on cosine similarity for alignment, which may not fully capture the complex relationship between medical images and reports. To address this gap, we introduce a novel approach called Cross-Attention Alignment for Radiology Zero-Shot Classification (CARZero). Our approach innovatively leverages cross-attention mechanisms to process image and report features, creating a Similarity Representation that more accurately reflects the intricate relationships in medical semantics. This representation is then linearly projected to form an image-text similarity matrix for cross-modality alignment. Additionally, recognizing the pivotal role of prompt selection in zero-shot learning, CARZero incorporates a Large Language Model-based prompt alignment strategy. This strategy standardizes diverse diagnostic expressions into a unified format for both training and inference phases, overcoming the challenges of manual prompt design. Our approach is simple yet effective, demonstrating state-of-the-art performance in zero-shot classification on five official chest radiograph diagnostic test sets, including remarkable results on datasets with long-tail distributions of rare diseases. This achievement is attributed to our new image-text alignment strategy, which effectively addresses the complex relationship between medical images and reports. Code and models are available at https://github.com/laihaoran/CARZero.

Tumor synthesis enables the creation of artificial tumor examples in medical images, facilitating the training of AI models for tumor detection and segmentation. However, success in tumor synthesis hinges on creating visually realistic tumors that are generalizable across multiple organs and, furthermore, the resulting AI models being capable of detecting real tumors in images sourced from different medical domains (e.g., hospitals). This paper made a progressive stride toward generalizable tumor synthesis by leveraging a critical observation: early-stage tumors ($<$ 2cm) tend to have similar imaging characteristics in computed tomography (CT), whether they originate in the liver, pancreas, or kidneys. We have ascertained that generative AI models, e.g., Diffusion Models, can create realistic tumors generalized to a range of organs even when trained on a limited number of tumor examples from just one organ. Moreover, we have shown that AI models trained on these synthetic tumors can be generalized to detect and segment real tumors from CT volumes, encompassing a broad spectrum of patient demographics, imaging protocols, and healthcare facilities. The code is attached as supplementary material and will be released.

Existing learning-based solutions to medical image segmentation have two important shortcomings. First, for most new segmentation task, a new model has to be trained or fine-tuned. This requires extensive resources and machine learning expertise, and is therefore often infeasible for medical researchers and clinicians. Second, most existing segmentation methods produce a single deterministic segmentation mask for a given image. In practice however, there is often considerable uncertainty about what constitutes the correct segmentation, and different expert annotators will often segment the same image differently. We tackle both of these problems with Tyche, a model that uses a context set to generate stochastic predictions for previously unseen tasks without the need to retrain. Tyche differs from other in-context segmentation methods in two important ways. (1) We introduce a novel convolution block architecture that enables interactions among predictions. (2) We introduce in-context test-time augmentation, a new mechanism to provide prediction stochasticity. When combined with appropriate model design and loss functions, Tyche can predict a set of plausible diverse segmentation candidates for new or unseen medical images and segmentation tasks without the need to retrain. Code available at: https://github.com/mariannerakic/tyche/

X-ray, known for its ability to reveal internal structures of objects, is expected to provide richer information for 3D reconstruction than visible light. Yet, existing neural radiance fields (NeRF) algorithms overlook this important nature of X-ray, leading to their limitations in capturing structural contents of imaged objects. In this paper, we propose a framework, Structure-Aware X-ray Neural Radiodensity Fields (SAX-NeRF), for sparse-view X-ray 3D reconstruction. Firstly, we design a Line Segment-based Transformer (Lineformer) as the backbone of SAX-NeRF. Linefomer captures internal structures of objects in 3D space by modeling the dependencies within each line segment of an X-ray. Secondly, we present a Masked Local-Global (MLG) ray sampling strategy to extract contextual and geometric information in 2D projection. Plus, we collect a larger-scale dataset X3D covering wider X-ray applications. Experiments on X3D show that SAX-NeRF surpasses previous NeRF-based methods by $\bf 12.56$ and $\bf 2.49~dB$ on novel view synthesis and CT reconstruction. Code, models, and data will be released.

Distribution shift widely exists in medical images acquired from different medical centres and poses a significant obstacle to deploying the pre-trained semantic segmentation model in real-world applications. Test-time adaptation has proven its effectiveness in tackling the cross-domain distribution shift during inference. However, most existing methods achieve adaptation by updating the pre-trained models, rendering them susceptible to error accumulation and catastrophic forgetting when encountering a series of distribution shifts (i.e., under the continual test-time adaptation setup). To overcome these challenges caused by updating the models, in this paper, we freeze the pre-trained model and propose the Visual Prompt-based Test-Time Adaptation (VPTTA) method to train a specific prompt for each test image to align the statistics in the batch normalization layers. Specifically, we present the low-frequency prompt, which is lightweight with only a few parameters and can be effectively trained in a single iteration. To enhance prompt initialization, we equip VPTTA with a memory bank to benefit the current prompt from previous ones. Additionally, we design a warm-up mechanism, which mixes source and target statistics to construct warm-up statistics, thereby facilitating the training process. Extensive experiments demonstrate the superiority of our VPTTA over other state-of-the-art methods on two medical image segmentation benchmark tasks. The code and weights of pre-trained source models are available at https://github.com/Chen-Ziyang/VPTTA.

A major focus of clinical imaging workflow is disease diagnosis and management, leading to medical imaging datasets strongly tied to specific clinical objectives. This scenario has led to the prevailing practice of developing task-specific segmentation models, without gaining insights from widespread imaging cohorts. Inspired by the training program of medical radiology residents, we propose a shift towards universal medical image segmentation, a paradigm aiming to leverage the diversity and commonality across clinical targets, body regions, and imaging modalities. Towards this goal, we develop Hermes, a novel context-prior learning approach to address the challenges of data heterogeneity and annotation differences in medical image segmentation. In a large collection of eleven diverse datasets (2,438 3D images) across five modalities (CT, PET, T1, T2 and cine MRI) and multiple body regions, we demonstrate the merit of the universal paradigm over the traditional paradigm on addressing multiple tasks within a single model. By exploiting the synergy across a spectrum of tasks, Hermes achieves state-of-the-art performance on all testing datasets and shows superior model scalability. Investigation on two additional datasets reveals Hermes' strong performance for transfer learning, incremental learning, and generalization to downstream tasks. Hermes's learned priors demonstrate an appealing trait to reflect the intricate relations among tasks and modalities, which well aligns with the established anatomical and imaging principles in radiology. The code and model will be made publicly available.

Cone beam computed tomography (CBCT) is an important imaging technology widely used in medical scenarios, such as diagnosis and preoperative planning. Using fewer projection views to reconstruct CT, also known as sparse-view reconstruction, can reduce ionizing radiation and further benefit interventional radiology. Compared with sparse-view reconstruction for traditional parallel/fan-beam CT, CBCT reconstruction is more challenging due to the increased dimensionality caused by the measurement process based on cone-shaped X-ray beams. As a 2D-to-3D reconstruction problem, although implicit neural representations have been introduced to enable efficient training, only local features are considered and different views are processed equally in previous works, resulting in spatial inconsistency and poor performance on complicated anatomies. To this end, we propose C$^\text{2}$RV by leveraging explicit multi-scale volumetric representations to enable cross-regional learning in the 3D space. Additionally, the scale-view cross-attention module is introduced to adaptively aggregate multi-scale and multi-view features. Extensive experiments demonstrate that our C$^\text{2}$RV achieves consistent and significant improvement over previous state-of-the-art methods on datasets with diverse anatomy.

Establishing dense anatomical correspondence across distinct imaging modalities is a foundational yet challenging procedure for numerous medical image analysis studies and image-guided radiotherapy. Existing multi-modality image registration algorithms rely on statistical-based similarity measures or local structural image representations. However, the former is sensitive to locally varying noise, while the latter is not discriminative enough to cope with complex anatomical structures in multimodal scans, causing ambiguity in determining the anatomical correspondence across scans with different modalities. In this paper, we propose a modality-agnostic structural representation learning method, which leverages Deep Neighbourhood Self-similarity (DNS) and anatomy-aware contrastive learning to learn discriminative and contrast-invariance deep structural image representations (DSIR) without the need for anatomical delineations or pre-aligned training images. We evaluate our method on multiphase CT, abdomen MR-CT, and brain MR T1w-T2w registration. Comprehensive results demonstrate that our method is superior to the conventional local structural representation and statistical-based similarity measures in terms of discriminability and accuracy.

Introducing interpretability and reasoning into Multiple Instance Learning (MIL) methods for Whole Slide Image (WSI) analysis is challenging, given the complexity of gigapixel slides. Traditionally, MIL interpretability is limited to identifying salient regions deemed pertinent for downstream tasks, offering little insight to the end-user (pathologist) regarding the rationale behind these selections. To address this, we propose Self-Interpretable MIL (SI-MIL), a method intrinsically designed for interpretability from the very outset. SI-MIL employs a deep MIL framework to guide an interpretable branch grounded on handcrafted pathological features, facilitating linear predictions. Beyond identifying salient regions, SI-MIL uniquely provides feature-level interpretations rooted in pathological insights for WSIs. Notably, SI-MIL, with its linear prediction constraints, challenges the prevalent myth of an inevitable trade-off between model interpretability and performance, demonstrating competitive results compared to state-of-the-art methods on WSI-level prediction tasks across three cancer types. In addition, we thoroughly benchmark the local- and global-interpretability of SI-MIL in terms of statistical analysis, a domain expert study, and desiderata of interpretability, namely, user-friendliness and faithfulness.

Radiologists highly desire fully automated versatile AI for medical imaging interpretation. However, the lack of extensively annotated large-scale multi-disease datasets has hindered the achievement of this goal. In this paper, we explore the feasibility of leveraging language as a naturally high-quality supervision for chest CT imaging. In light of the limited availability of image-report pairs, we bootstrap the understanding of 3D chest CT images by distilling chest-related diagnostic knowledge from an extensively pre-trained 2D X-ray expert model. Specifically, we propose a language-guided retrieval method to match each 3D CT image with its semantically closest 2D X-ray image, and perform pair-wise and semantic relation knowledge distillation. Subsequently, we use contrastive learning to align images and reports within the same patient while distinguishing them from the other patients. However, the challenge arises when patients have similar semantic diagnoses, such as healthy patients, potentially confusing if treated as negatives. We introduce a robust contrastive learning that identifies and corrects these false negatives. We train our model with over 12K pairs of chest CT images and radiology reports. Extensive experiments across multiple scenarios, including zero-shot learning, report generation, and fine-tuning processes, demonstrate the model’s feasibility in interpreting chest CT images.

Multiple instance learning (MIL)-based framework has become the mainstream for processing the whole slide image (WSI) with giga-pixel size and hierarchical image context in digital pathology. However, these methods heavily depend on a substantial number of bag-level labels and solely learn from the original slides, which are easily affected by variations in data distribution. Recently, vision language model (VLM)-based methods introduced the language prior by pre-training on large-scale pathological image-text pairs. However, the previous text prompt lacks the consideration of pathological prior knowledge, therefore does not substantially boost the model's performance. Moreover, the collection of such pairs and the pre-training process are very time-consuming and source-intensive. To solve the above problems, we propose a dual-scale vision-language multiple instance learning (ViLa-MIL) framework for whole slide image classification. Specifically, we propose a dual-scale visual descriptive text prompt based on the frozen large language model (LLM) to boost the performance of VLM effectively. To transfer the VLM to process WSI efficiently, for the image branch, we propose a prototype-guided patch decoder to aggregate the patch features progressively by grouping similar patches into the same prototype; for the text branch, we introduce a context-guided text decoder to enhance the text features by incorporating the multi-granular image contexts. Extensive studies on three multi-cancer and multi-center subtyping datasets demonstrate the superiority of ViLa-MIL.

Recently, virtual staining technology has greatly promoted the advancement of histopathology.Despite the practical successes achieved, the outstanding performance of most virtual staining methods relies on hard-to-obtain paired images in training.In this paper, we propose a method for virtual immunohistochemistry (IHC) staining, named confusion-GAN, which does not require paired images and can achieve comparable performance to supervised algorithms.Specifically, we propose a multi-branch discriminator, which judges if the features of generated images can be embedded into the feature pool of target domain images, to improve the visual quality of generated images.Meanwhile, we also propose a novel patch-level pathological information extractor, which is assisted by multiple instance learning, to ensure pathological consistency during virtual staining.Extensive experiments were conducted on three types of IHC images, including a high-resolution hepatocellular carcinoma immunohistochemical dataset proposed by us. The results demonstrated that our proposed confusion-GAN can generate highly realistic images that are capable of deceiving even experienced pathologists.Furthermore, compared to using H\&E images directly, the downstream diagnosis achieved higher accuracy when using images generated by confusion-GAN.Our dataset and codes will be available at https://github.com/jiahanli2022/confusion-GAN.

Humans effortlessly interpret images by parsing them into part-whole hierarchies; deep learning excels in learning multi-level feature spaces, but they often lack explicit coding of part-whole relations, a prominent property of medical imaging. To overcome this limitation, we introduce Adam–v2, a new self-supervised learning framework extending Adam [69] by explicitly incorporating part-whole hierarchies into its learning objectives through three key branches: (1) Localizability, acquiring discriminative representations to distinguish different anatomical patterns; (2) Composability, learning each anatomical structure in a parts-to-whole manner; and (3) Decomposability, comprehending each anatomical structure in a whole-to-parts manner. Experimental results across 10 tasks, compared to 11 baselines in zero-shot, few-shot transfer, and full fine-tuning settings, showcase Adam–v2’s superior performance over large-scale medical models and existing SSL methods across diverse downstream tasks. The higher generality and robustness of Adam–v2’s representations originate from its explicit construction of hierarchies for distinct anatomical structures from unlabeled medical images. Adam–v2 preserves a semantic balance of anatomical diversity and harmony in its embedding, yielding representations that are both generic and semantically meaningful, yet overlooked in existing SSL methods. All code and pretrained models are available at GitHub.com/JLiangLab/Eden.

The increasing prevalence of non-alcoholic fatty liver disease (NAFLD) has caused public concern in recent years. Fibrosis staging from liver biopsy images plays a key role in demonstrating the histological progression of NAFLD. Fibrosis mainly involves the deposition of fibers around vessels. Current deep learning-based fibrosis staging methods learn spatial relationships between tissue patches but do not explicitly consider the relationships between vessels and fibers, leading to limited performance and poor interpretability. In this paper, we propose an e$\textbf{X}$plicit vessel-fiber modeling method for $\textbf{Fibrosis}$ staging from liver biopsy images, namely XFibrosis. Specifically, we transform vessels and fibers into graph-structured representations, where their micro-structures are depicted by vessel-induced primal graphs and fiber-induced dual graphs, respectively. Moreover, the fiber-induced dual graphs also represent the connectivity information between vessels caused by fiber deposition. A primal-dual graph convolution module is designed to facilitate the learning of spatial relationships between vessels and fibers, allowing for the joint exploration and interaction of their micro-structures. Experiments conducted on two datasets have shown that explicitly modeling the relationship between vessels and fibers leads to improved fibrosis staging and enhanced interpretability.

The rapid increase in cases of non-alcoholic fatty liver disease (NAFLD) in recent years has raised significant public concern. Accurately identifying tissue alteration regions is crucial for the diagnosis of NAFLD, but this task presents challenges in pathology image analysis, particularly with small-scale datasets. Recently, the paradigm shift from full fine-tuning to prompting in adapting vision foundation models has offered a new perspective for small-scale data analysis. However, existing prompting methods based on task-agnostic prompts are mainly developed for generic image recognition, which fall short in providing instructive cues for complex pathology images. In this paper, we propose ${Q}$uantitative ${A}$ttribute-based ${P}$rompting (QAP), a novel prompting method specifically for liver pathology image analysis. QAP is based on two quantitative attributes, namely K-function-based spatial attributes and histogram-based morphological attributes, which are aimed for quantitative assessment of tissue states. Moreover, a conditional prompt generator is designed to turn these instance-specific attributes into visual prompts. Extensive experiments on three diverse tasks demonstrate that our task-specific prompting method achieves better diagnostic performance as well as better interpretability.

Large foundation models, known for their strong zero-shot generalization, have excelled in visual and language applications. However, applying them to medical image segmentation, a domain with diverse imaging types and target labels, remains an open challenge. Current approaches, such as adapting interactive segmentation models like Segment Anything Model (SAM), require user prompts for each sample during inference. Alternatively, transfer learning methods like few/one-shot models demand labeled samples, leading to high costs. This paper introduces a new paradigm toward the universal medical image segmentation, termed 'One-Prompt Segmentation.' One-Prompt Segmentation combines the strengths of one-shot and interactive methods. In the inference stage, with just one prompted sample, it can adeptly handle the unseen task in a single forward pass. We train One-Prompt Model on 64 open-source medical datasets, accompanied by the collection of over 3,000 clinician-labeled prompts. Tested on 14 previously unseen tasks, the One-Prompt Model showcases superior zero-shot segmentation capabilities, outperforming a wide range of related methods. The code and annotated data will be publicly released.

As the mainstream technique for capturing images of biological specimens at nanometer resolution, electron microscopy (EM) is extremely time-consuming for scanning wide field-of-view (FOV) specimens. In this paper, we investigate a challenging task of large-factor EM image super-resolution (EMSR), which holds great promise for reducing scanning time, relaxing acquisition conditions, and expanding imaging FOV. By exploiting the repetitive structures and volumetric coherence of EM images, we propose the first generative learning-based framework for large-factor EMSR. Specifically, motivated by the predictability of repetitive structures and textures in EM images, we first learn a discrete codebook in the latent space to represent high-resolution (HR) cell-specific priors and a latent vector indexer to map low-resolution (LR) EM images to their corresponding latent vectors in a generative manner. By incorporating the generative cell-specific priors from HR EM images through a multi-scale prior fusion module, we then deploy multi-image feature alignment and fusion to further exploit the inter-section coherence in the volumetric EM data. Extensive experiments demonstrate that our proposed framework outperforms advanced single-image and video super-resolution methods for $8\times$ and $16\times$ EMSR (\textit{i.e.}, with 64 times and 256 times less data acquired, respectively), achieving superior visual reconstruction quality and downstream segmentation accuracy on benchmark EM datasets. Code is available at \url{https://github.com/jtshou/GPEMSR}.

Histopathological whole slide images (WSIs) classification has become a foundation task in medical microscopic imaging processing. Prevailing approaches involve learning WSIs as instance-bag representations, emphasizing significant instances but struggling to capture the interactions between instances. Additionally, conventional graph representation methods utilize explicit spatial positions to construct topological structures but restrict the flexible interaction capabilities between instances at arbitrary locations, particularly when spatially distant. In response, we propose a novel dynamic graph representation algorithm that conceptualizes WSIs as a form of the knowledge graph structure. Specifically, we dynamically construct neighbors and directed edge embeddings based on the head and tail relationships between instances. Then, we devise a knowledge-aware attention mechanism that can update the head node features by learning the joint attention score of each neighbor and edge. Finally, we obtain a graph-level embedding through the global pooling process of the updated head, serving as an implicit representation for the WSI classification. Our end-to-end graph representation learning approach has outperformed the state-of-the-art WSI analysis methods on three TCGA benchmark datasets and in-house test sets. Our code is available at https://github.com/WonderLandxD/WiKG.

Brain decoding, a pivotal field in neuroscience, aims to reconstruct stimuli from acquired brain signals, primarily utilizing functional magnetic resonance imaging (fMRI). Currently, brain decoding is confined to a per-subject-per-model paradigm, limiting its applicability to the same individual for whom the decoding model is trained. This constraint stems from three key challenges: 1) the inherent variability in input dimensions across subjects due to differences in brain size; 2) the unique intrinsic neural patterns, influencing how different individuals perceive and process sensory information; 3) limited data availability for new subjects in real-world scenarios hampers the performance of decoding models.In this paper, we present a novel approach, MindBridge, that achieves cross-subject brain decoding by employing only one model. Our proposed framework establishes a generic paradigm capable of addressing these challenges by introducing biological-inspired aggregation function and novel cyclic fMRI reconstruction mechanism for subject-invariant representation learning. Notably, by cycle reconstruction of fMRI, MindBridge can enable novel fMRI synthesis, which also can serve as pseudo data augmentation. Within the framework, we also devise a novel reset-tuning method for adapting a pretrained model to a new subject. Experimental results demonstrate MindBridge's ability to reconstruct images for multiple subjects, which is competitive with dedicated subject-specific models. Furthermore, with limited data for a new subject, we achieve a high level of decoding accuracy, surpassing that of subject-specific models. This advancement in cross-subject brain decoding suggests promising directions for wider applications in neuroscience and indicates potential for more efficient utilization of limited fMRI data in real-world scenarios. Project page: https://littlepure2333.github.io/MindBridge

Multiple instance learning (MIL) is the most widely used framework in computational pathology, encompassing sub-typing, diagnosis, prognosis, and more. However, the existing MIL paradigm typically requires an offline instance feature extractor, such as a pre-trained ResNet or a foundation model. This approach lacks the capability for feature fine-tuning within the specific downstream tasks, limiting its adaptability and performance. To address this issue, we propose a Re-embedded Regional Transformer (R$^2$T) for re-embedding the instance features online, which captures fine-grained local features and establishes connections across different regions. Unlike existing works that focus on pre-training powerful feature extractor or designing sophisticated instance aggregator, R$^2$T is tailored to re-embed instance features online. It serves as a portable module that can seamlessly integrate into mainstream MIL models. Extensive experimental results on common computational pathology tasks validate that: 1) feature re-embedding improves the performance of MIL models based on ResNet-50 features to the level of foundation model features, and further enhances the performance of foundation model features; 2) the R$^2$T can introduce more significant performance improvements to various MIL models; 3) R$^2$T-MIL, as an R$^2$T-enhanced AB-MIL, outperforms other latest methods by a large margin. The code is available at: https://github.com/DearCaat/RRT-MIL.

4D medical images, which represent 3D images with temporal information, are crucial in clinical practice for capturing dynamic changes and monitoring long-term disease progression. However, acquiring 4D medical images poses challenges due to factors such as radiation exposure and imaging duration, necessitating a balance between achieving high temporal resolution and minimizing adverse effects. Given these circumstances, not only is data acquisition challenging, but increasing the frame rate for each dataset also proves difficult. To address this challenge, this paper proposes a simple yet effective Unsupervised Volumetric Interpolation framework, UVI-Net. This framework facilitates temporal interpolation without the need for any intermediate frames, distinguishing it from the majority of other existing unsupervised methods. Experiments on benchmark datasets demonstrate significant improvements across diverse evaluation metrics compared to unsupervised and supervised baselines. Remarkably, our approach achieves this superior performance even when trained with a dataset as small as one, highlighting its exceptional robustness and efficiency in scenarios with sparse supervision. This positions UVI-Net as a compelling alternative for 4D medical imaging, particularly in settings where data availability is limited. The source code will be made available online.

Recently, diffusion models (DM) have been introduced in magnetic resonance imaging (MRI) super-resolution (SR) reconstruction, exhibiting impressive performance, particularly with regard to detailed reconstruction. However, the current DM-based SR reconstruction methods still face the following issues: (1) They require a large number of iterations to reconstruct the final image, which is inefficient and consumes a significant amount of computational resources. (2) The results reconstructed by these methods are often misaligned with the real high-resolution images, leading to remarkable distortion in the reconstructed MR images. To address the aforementioned issues, we propose an efficient diffusion model for multi-contrast MRI SR, named as McDiff. Specifically, we apply DM in a highly compact low-dimensional latent space to generate prior knowledge with high-frequency detail information. The highly compact latent space ensures that DM requires only a few simple iterations to produce accurate prior knowledge. In addition, we design the Prior-Guide Large Window Transformer (PLWformer) as the decoder for DM, which can extend the receptive field while fully utilizing the prior knowledge generated by DM to ensure that the reconstructed MR image remains undistorted. Extensive experiments on public and clinical datasets demonstrate that our McDiff outperforms state-of-the-art methods.

Recent advancements in large-scale visual-language pre-trained models have led to significant progress in zero-/few-shot anomaly detection within natural image domains. However, the substantial domain divergence between natural and medical images limits the effectiveness of these methodologies in medical anomaly detection. This paper introduces a novel lightweight multi-level adaptation and comparison framework to repurpose the CLIP model for medical anomaly detection. Our approach integrates multiple residual adapters into the pre-trained visual encoder, enabling a stepwise enhancement of visual features across different levels. This multi-level adaptation is guided by multi-level, pixel-wise visual-language feature alignment loss functions, which recalibrate the model’s focus from object semantics in natural imagery to anomaly identification in medical images. The adapted features exhibit improved generalization across various medical data types, even in zero-shot scenarios where the model encounters unseen medical modalities and anatomical regions during training. Our experiments on medical anomaly detection benchmarks demonstrate that our method significantly surpasses current state-of-the-art models, with an average AUC improvement of 6.24% and 7.33% for anomaly classification, 2.03\% and 2.37\% for anomaly segmentation, under the zero-shot and few-shot settings, respectively.

The long-tailed distribution problem of medical image analysis reflects a high prevalence of common conditions and a low prevalence of rare ones, which poses a significant challenge in developing a unified model capable of identifying rare or novel tumor categories not encountered during training. In this paper, we propose a new Zero-shot Pan-Tumor segmentation framework (ZePT) based on query-disentangling and self-prompting to segment unseen tumor categories beyond the training set. ZePT disentangles the object queries into two subsets and trains them in two stages. Initially, it learns a set of fundamental queries for organ segmentation through an object-aware feature grouping strategy, which gathers organ-level visual features. Subsequently, it refines the other set of advanced queries that focus on the auto-generated visual prompts for unseen tumor segmentation. Moreover, we introduce query-knowledge alignment at the feature level to enhance each query's discriminative representation and generalizability. Extensive experiments on various tumor segmentation tasks demonstrate the performance superiority of ZePT, which surpasses the previous counterparts and demonstrates the promising ability for zero-shot tumor segmentation in real-world settings.

We propose a novel echocardiographical video segmentation model by adapting SAM to medical videos to address some long-standing challenges in ultrasound video segmentation, including (1) massive speckle noise and artifacts, (2) extremely ambiguous boundaries, and (3) large variations of targeting objects across frames. The core technique of our model is a temporal-aware and noise-resilient prompting scheme. Specifically, we employ a space-time memory that contains both spatial and temporal information to prompt the segmentation of current frame, and thus we call the proposed model as MemSAM. In prompting, the memory carrying temporal cues sequentially prompt the video segmentation frame by frame. Meanwhile, as the memory prompt propagates high-level features, it avoids the issue of misidentification caused by mask propagation and improves representation consistency. To address the challenge of speckle noise, we further propose a memory reinforcement mechanism, which leverages predicted masks to improve the quality of the memory before storing it. We extensively evaluate our method on two public datasets and demonstrate state-of-the-art performance compared to existing models. Particularly, our model achieves comparable performance with fully supervised approaches with limited annotations. Codes are available at https://github.com/dengxl0520/MemSAM.

Whole Slide Image (WSI) classification is often formulated as a Multiple Instance Learning (MIL) problem. Recently, Vision-Language Models (VLMs) have demonstrated remarkable performance in WSI classification. However, existing methods leverage coarse-grained pathogenetic descriptions for visual representation supervision, which are insufficient to capture the complex visual appearance of pathogenetic images, hindering the generalizability of models on diverse downstream tasks. Additionally, processing high-resolution WSIs can be computationally expensive. In this paper, we propose a novel ``Fine-grained Visual-Semantic Interaction" (FiVE) framework for WSI classification. It is designed to enhance the model's generalizability by leveraging the interaction between localized visual patterns and fine-grained pathological semantics. Specifically, with meticulously designed queries, we start by utilizing a large language model to extract fine-grained pathological descriptions from various non-standardized raw reports. The output descriptions are then reconstructed into fine-grained labels used for training. By introducing a Task-specific Fine-grained Semantics (TFS) module, we enable prompts to capture crucial visual information in WSIs, which enhances representation learning and augments generalization capabilities significantly. Furthermore, given that pathological visual patterns are redundantly distributed across tissue slices, we sample a subset of visual instances during training. Our method demonstrates robust generalizability and strong transferability, dominantly outperforming the counterparts on the TCGA Lung Cancer dataset with at least 9.19\% higher accuracy in few-shot experiments. The code is available at: https://github.com/ls1rius/WSI_FiVE.

We present a novel semantic segmentation approach for incremental nuclei segmentation from histopathological images, which is a very challenging task as we have to incrementally optimize existing models to make them perform well in both old and new classes without using training samples of old classes. Yet, it is an indispensable component of computer-aided diagnosis systems. The proposed approach has two key techniques. First, we propose a new future-class awareness mechanism by separating some potential regions for future classes from background based on their similarities to both old and new classes in the representation space. With this mechanism, we can not only reserve more parameter space for future updates but also enhance the representation capability of learned features. We further propose an innovative compatibility-inspired distillation scheme to make our model take full advantage of the knowledge learned by the old model. We conducted extensive experiments on two famous histopathological datasets and the results demonstrate the proposed approach achieves much better performance than state-of-the-art approaches. The code is available at https://github.com/why19991/InSeg.

We present a novel semi-supervised framework for breast ultrasound (BUS) image segmentation, which is a very challenging task owing to (1) large scale and shape variations of breast lesions and (2) extremely ambiguous boundaries caused by massive speckle noise and artifacts in BUS images. While existing models achieved certain progress in this task, we believe the main bottleneck nowadays for further improvement is that we still cannot deal with hard cases well. Our framework aims to break through this bottleneck, which includes two innovative components: an adaptive patch augmentation scheme and a hard-patch contrastive learning module. We first identify hard patches by computing the average entropy of each patch and then shield hard patches to prevent them from being cropped out while performing random patch cutmix. Such a scheme is able to prevent hard regions from being inadequately trained under strong augmentation. We further develop a new hard-patch contrastive learning algorithm to direct model attention to hard regions by applying extra contrast to pixels in hard patches, further improving segmentation performance on hard cases. We demonstrate the superiority of our framework to state-of-the-art approaches on two famous BUS datasets, achieving better performance under different labeling conditions. The code is available at https://github.com/jjjsyyy/PH-Net.

Annotating lots of 3D medical images for training segmentation models is time-consuming.The goal of weakly supervised semantic segmentation is to train segmentation models without using any ground truth segmentation masks. Our work addresses the case where only image-level categorical labels, indicating the presence or absence of a particular region of interest (such as tumours or lesions), are available.Most existing methods rely on class activation mapping (CAM). We propose a novel approach, ToNNO, which is based on the Tomographic reconstruction of a Neural Network's Output. Our technique extracts stacks of slices with different angles from the input 3D volume, feeds these slices to a 2D encoder, and applies the inverse Radon transform in order to reconstruct a 3D heatmap of the encoder’s predictions. This generic method allows to perform dense prediction tasks on 3D volumes using any 2D image encoder. We apply it to weakly supervised medical image segmentation by training the 2D encoder to output high values for slices containing the regions of interest. We test it on four large scale medical image datasets and outperform 2D CAM methods. We also propose to combine tomographic reconstruction with CAM methods in order to obtain even better results.

Federated learning facilitates the collaborative learning of a global model across multiple distributed medical institutions without centralizing data. Nevertheless, the expensive cost of annotation on local clients remains an obstacle to effectively utilizing local data. To mitigate this issue, federated active learning methods suggest leveraging local and global model predictions to select a relatively small amount of informative local data for annotation. However, existing methods mainly focus on all local data sampled from the same domain, making them unreliable in realistic medical scenarios with domain shifts among different clients. In this paper, we make the first attempt to assess the informativeness of local data derived from diverse domains and propose a novel methodology termed Federated Evidential Active Learning (FEAL) to calibrate the data evaluation under domain shift. Specifically, we introduce a Dirichlet prior distribution in both local and global models to treat the prediction as a distribution over the probability simplex and capture both aleatoric and epistemic uncertainties by using the Dirichlet-based evidential model. Then we employ the epistemic uncertainty to calibrate the aleatoric uncertainty. Afterward, we design a diversity relaxation strategy to reduce data redundancy and maintain data diversity. Extensive experiments and analyses are conducted to show the superiority of FEAL over the state-of-the-art active learning methods and the efficiency of FEAL under the federated active learning framework.

This paper proposes Comprehensive Pathology Language Image Pre-training (CPLIP), a new unsupervised technique designed to enhance the alignment of images and text in histopathology for tasks such as classification and segmentation. This methodology enriches vision-language models by leveraging extensive data without needing ground truth annotations. CPLIP involves constructing a pathology-specific vocabulary, generating textual descriptions for images using language models, and retrieving relevant images for each text snippet via a pre-trained model. The model is then fine-tuned using a many-to-many contrastive learning method to align complex interrelated concepts across both modalities. Evaluated across multiple histopathology tasks, CPLIP shows notable improvements in zero-shot learning scenarios, outperforming existing methods in both interpretability and robustness and setting a higher benchmark for the application of vision-language models in the field. To encourage further research and replication, the code for CPLIP is available on GitHub at xxx.

Self-supervised learning (SSL) has been successful in building patch embeddings of small histology images (e.g., 224 x 224 pixels), but scaling these models to learn slide embeddings from the entirety of giga-pixel whole-slide images (WSIs) remains challenging. Here, we leverage complementary information from gene expression profiles to guide slide representation learning using multimodal pre-training. Expression profiles constitute highly detailed molecular descriptions of a tissue that we hypothesize offer a strong task-agnostic training signal for learning slide embeddings. Our slide and expression (S+E) pre-training strategy, called TANGLE, employs modality-specific encoders, the outputs of which are aligned via contrastive learning. TANGLE was pre-trained on samples from three different organs: liver (n=6,597 S+E pairs), breast (n=1,020), and lung (n=1,012) from two different species (Homo sapiens and Rattus norvegicus). Across three independent test datasets consisting of 1,265 breast WSIs, 1,946 lung WSIs, and 4,584 liver WSIs, TANGLE shows significantly better few-shot performance compared to supervised and SSL baselines. When assessed using prototype-based classification and slide retrieval, TANGLE also shows a substantial performance improvement over all baselines. Code will be made available upon acceptance.

Volumetric optical microscopy using non-diffracting beams, offers rapid imaging of 3D volumes by projecting them axially to 2D images, but it often falls short in providing depth information. To address this limitation, we introduce MicroDiffusion, a pioneering tool designed for high-quality, depth-resolved 3D volume reconstruction from a limited set of 2D microscopy projections. Existing 3D reconstruction methods, such as Implicit Neural Representation (INR) models, often result in incomplete and noisy outputs. In contrast, Denoising Diffusion Probabilistic Models (DDPM) excel at capturing fine-grained details. Our method merges the strengths of INR’s ability to maintain structural 3D coherence with DDPM’s proficiency in enhancing details.Initially, we pretrain an INR model that transforms the 2D axially-projected images into a preliminary 3D volume. Then, the pretrained INR serves as a global prior, directing DDPM's generative process through linear interpolation between INR outputs and noise inputs. This strategy effectively enriches the diffusion process with structured 3D information while simultaneously enhancing detail and minimizing noise in localized 2D images. Furthermore, by conditioning the diffusion model on the closest 2D image, MicroDiffusion substantially enhances the fidelity of the resulting 3D reconstructions. MicroDiffusion enables depth-resolved volumetric microscopy by delivering high-quality 3D reconstructions that are sharper than those produced by INR models and more coherent than standard DDPM outputs.Extensive results from three microscopy datasets demonstrate MicroDiffusion's superiority in producing 3D reconstructions with enhanced image quality, structural coherence, and fidelity, compared to traditional INR and diffusion models.

Annotation ambiguity due to inherent data uncertainties such as blurred boundaries in medical scans and different observer expertise and preferences has become a major obstacle for training deep-learning based medical image segmentation models. To address it, the common practice is to gather multiple annotations from different experts, leading to the setting of multi-rater medical image segmentation. Existing works aim to either merge different annotations into the "groundtruth" that is often unattainable in numerous medical contexts, or generate diverse results, or produce personalized results corresponding to individual expert raters. Here, we bring up a more ambitious goal for multi-rater medical image segmentation, i.e., obtaining both diversified and personalized results. Specifically, we propose a two-stage framework named D-Persona (first Diversification and then Personalization). In Stage I, we exploit multiple given annotations to train a Probabilistic U-Net model, with a bound-constrained loss to improve the prediction diversity. In this way, a common latent space is constructed in Stage I, where different latent codes denote diversified expert opinions. Then, in Stage II, we design multiple attention-based projection heads to adaptively query the corresponding expert prompts from the shared latent space, and then perform the personalized medical image segmentation. We evaluated the proposed model on our in-house Nasopharyngeal Carcinoma dataset and the public lung nodule dataset (i.e., LIDC-IDRI). Extensive experiments demonstrated our D-Persona can provide diversified and personalized results at the same time, achieving new SOTA performance for multi-rater medical image segmentation. Our code will be released publicly.

Generalizability in deep neural networks plays a pivotal role in medical image segmentation. However, deep learning-based medical image analyses tend to overlook the importance of frequency variance, which is critical element for achieving a model that is both modality-agnostic and domain-generalizable. Additionally, various models fail to account for the potential information loss that can arise from multi-task learning under deep supervision, a factor that can impair the model’s representation ability. To address these challenges, we propose a Modality-agnostic Domain Generalizable Network (MADGNet) for medical image segmentation, which comprises two key components: a Multi-Frequency in Multi-Scale Attention (MFMSA) block and Ensemble Sub-Decoding Module (E-SDM). The MFMSA block refines the process of spatial feature extraction, particularly in capturing boundary features, by incorporating multi-frequency and multi-scale features, thereby offering informative cues for tissue outline and anatomical structures. Moreover, we propose E-SDM to mitigate information loss in multi-task learning with deep supervision, especially during substantial upsampling from low resolution. We evaluate the segmentation performance of MADGNet across six modalities and fifteen datasets. Through extensive experiments, we demonstrate that MADGNet consistently outperforms state-of-the-art models across various modalities, showcasing superior segmentation performance. This affirms MADGNet as a robust solution for medical image segmentation that excels in diverse imaging scenarios.

Medical vision language pre-training (VLP) has emerged as a frontier of research, enabling zero-shot pathological recognition by comparing the query image with the textual descriptions for each disease. Due to the complex semantics of biomedical texts, current methods struggle to align medical images with key pathological findings in unstructured reports. This leads to the misalignment with the target disease's textual representation. In this paper, we introduce a novel VLP framework designed to dissect disease descriptions into their fundamental aspects, leveraging prior knowledge about the visual manifestations of pathologies. This is achieved by consulting a large language model and medical experts. Integrating a Transformer module, our approach aligns an input image with the diverse elements of a disease, generating aspect-centric image representations. By consolidating the matches from each aspect, we improve the compatibility between an image and its associated disease. Additionally, capitalizing on the aspect-oriented representations, we present a dual-head Transformer tailored to process known and unknown diseases, optimizing the comprehensive detection efficacy. Conducting experiments on seven downstream datasets, ours improves the accuracy of recent methods by up to 8.56% and 17.0% for seen and unseen categories, respectively. Our code is released at https://github.com/HieuPhan33/MAVL.

Medical generative models, acknowledged for their high-quality sample generation ability, have accelerated the fast growth of medical applications. However, recent works concentrate on separate medical generation models for distinct medical tasks and are restricted to inadequate medical multi-modal knowledge, constraining medical comprehensive diagnosis.In this paper, we propose MedM2G, a Medical Multi-Modal Generative framework, with the key innovation to align, extract and generate medical multi-modal within a unified model. Extending beyond single or two medical modalities, we efficiently align medical multi-modal through the central alignment approach in the unified space. Significantly, our framework extracts valuable clinical knowledge by preserving the medical visual invariant of each imaging modal, thereby enhancing specific medical information for multi-modal generation. By conditioning the adaptive cross-guided parameters into the multi-flow diffusion framework, our model promotes flexible interactions among medical multi-modal for generation. MedM2G is the first medical generative model that unifies medical generation tasks of text-to-image, image-to-text, and unified generation of medical modalities (CT, MRI, X-ray). It performs 5 medical generation tasks across 10 datasets, consistently outperforming various state-of-the-art works.

This paper introduces a novel top-down representation approach for deformable image registration, which estimates the deformation field by capturing various short- and long-range flow features at different scale levels. As a Hierarchical Vision Transformer (H-ViT), we propose a dual self-attention and cross-attention mechanism that uses high-level features in the deformation field to represent low-level ones, enabling information streams in the deformation field across all voxel patch embeddings irrespective of their spatial proximity. Since high-level features contain abstract flow patterns, such patterns are expected to positively contribute to the representation of the deformation field in lower scales. When the self-attention module utilizes within-scale short-range patterns for representation, the cross-attention modules dynamically look for the key tokens across different scales to further interact with the local query voxel patches. Our method shows superior accuracy and visual quality in deformable image registration over the state-of-the-art in five publicly available datasets, highlighting a substantial enhancement in the performance of medical imaging registration. The code and pre-trained models are available at https://github.com/---.

Machine learning holds tremendous promise for transforming the fundamental practice of scientific discovery by virtue of its data-driven nature. With the ever-increasing stream of research data collection, it would be appealing to autonomously explore patterns and insights from observational data for discovering novel classes of phenotypes and concepts. However, in the biomedical domain, there are several challenges inherently presented in the cumulated data which hamper the progress of novel class discovery. The non-i.i.d. data distribution accompanied by the severe imbalance among different groups of classes essentially leads to ambiguous and biased semantic representations. In this work, we present a geometry-constrained probabilistic modeling treatment to resolve the identified issues. First, we propose to parameterize the approximated posterior of instance embedding as a marginal von Mises-Fisher distribution to account for the interference of distributional latent bias. Then, we incorporate a suite of critical geometric properties to impose proper constraints on the layout of constructed embedding space, which in turn minimizes the uncontrollable risk for unknown class learning and structuring. Furthermore, a spectral graph-theoretic method is devised to estimate the number of potential novel classes. It inherits two intriguing merits compared to existent approaches, namely high computational efficiency and flexibility for taxonomy-adaptive estimation. Extensive experiments across various biomedical scenarios substantiate the effectiveness and general applicability of our method.

In magnetic resonance imaging (MRI), slice-to-volume reconstruction (SVR) refers to computational reconstruction of an unknown 3D magnetic resonance volume from stacks of 2D slices corrupted by motion. While promising, current SVR methods require multiple slice stacks for accurate 3D reconstruction, leading to long scans and limiting their use in time-sensitive applications such as fetal fMRI. Here, we propose a SVR method that overcomes the shortcomings of previous work and produces state-of-the-art reconstructions in the presence of extreme inter-slice motion. Inspired by the recent success of single-view depth estimation methods, we formulate SVR as a single-stack motion estimation task and train a fully convolutional network to predict a motion stack for a given slice stack, producing a 3D reconstruction as a byproduct of the predicted motion. Extensive experiments on the SVR of adult and fetal brains demonstrate that our fully convolutional method is twice as accurate as previous SVR methods. Our code is available at github.com/seannz/svr.

Deep learning-based image registration (DLIR) methods have achieved remarkable success in deformable image registration. We observe that iterative inference can exploit the well-trained registration network to the fullest extent. In this work, we propose a novel Iterative Inference Residual Pyramid Network (IIRP-Net) to enhance registration performance without any additional training costs. In IIRP-Net, we construct a streamlined pyramid registration network consisting of a feature extractor and residual flow estimators (RP-Net) to achieve generalized capabilities in feature extraction and registration. Then, in the inference phase, IIRP-Net employs an iterative inference strategy to enhance RP-Net by iteratively reutilizing residual flow estimators from coarse to fine. The number of iterations is adaptively determined by the proposed IterStop mechanism. We conduct extensive experiments on the FLARE and Mindboggle datasets and the results verify the effectiveness of the proposed method, outperforming state-of-the-art deformable image registration methods. Our code is available at https://github.com/Torbjorn1997/IIRP-Net.

Unlike color photography images, which are consistently encoded into RGB channels, biological images encompass various modalities, where the type of microscopy and the meaning of each channel varies with each experiment. Importantly, the number of channels can range from one to a dozen and their correlation is often comparatively much lower than RGB, as each of them brings specific information content. This aspect is largely overlooked by methods designed out of the bioimage field, and current solutions mostly focus on intra-channel spatial attention, often ignoring the relationship between channels, yet crucial in most biological applications. Importantly, the variable channel type and count prevent the projection of several experiments to a unified representation for large scale pre-training. In this study, we propose ChAda-ViT, a novel Channel Adaptive Vision Transformer architecture employing an Inter-Channel Attention mechanism on images with an arbitrary number, order and type of channels. We also introduce IDRCell100k, a bioimage dataset with a rich set of 79 experiments covering 7 microscope modalities, with a multitude of channel types, and channel counts varying from 1 to 10 per experiment. Our proposed architecture, trained in a self-supervised manner, outperforms existing approaches in several biologically relevant downstream tasks. Additionally, it can be used to bridge the gap for the first time between assays with different microscopes, channel numbers or types by embedding various image and experimental modalities into a unified biological image representation. The latter should facilitate interdisciplinary studies and pave the way for better adoption of deep learning in biological image-based analyses. Code and data will be available upon acceptance.

Representation learning of whole-slide images (WSIs) has been mainly based on weak supervision with Multiple Instance Learning (MIL). However, the resulting slide embedding is highly tailored to the clinical task at hand, which limits model expressivity and generalization, especially in low-data regimes. Instead, we hypothesize that morphological redundancy can be leveraged to build a task-agnostic slide representation in an unsupervised fashion. To this end, we introduce PANTHER, a novel method based on the Gaussian mixture model that summarizes the set of WSI patches with morphological prototypes. Specifically, each patch is assumed to have been generated from a mixture distribution, where each mixture component represents a morphological exemplar. We then construct a compact slide representation from the estimated mixture parameters that can be readily used for a range of downstream tasks.By performing an extensive evaluation of PANTHER on survival and subtyping tasks using 13 datasets, we show that 1) PANTHER outperforms or is on par with supervised MIL and 2) the analysis of morphological prototypes brings new qualitative and quantitative insights into model interpretability. The code will be released upon acceptance.

Integrating whole-slide images (WSIs) and bulk transcriptomics for predicting patient survival can improve our understanding of patient prognosis. However, this multimodal task is particularly challenging due to the different nature of these data: WSIs represent a very high-dimensional spatial description of a tumor, while bulk transcriptomics represent a global description of gene expression levels within that tumor. In this context, our work aims to address two key challenges: (1) how can we tokenize transcriptomics in a semantically meaningful and interpretable way?, and (2) how can we capture dense multimodal interactions between these two modalities? Here, we propose to learn biological pathway tokens from transcriptomics that can encode specific cellular functions. Together with histology patch tokens that encode the slide morphology, we argue that they form appropriate reasoning units for interpretability. We fuse both modalities using a memory-efficient multimodal Transformer that can model interactions between pathway and histology patch tokens. Our model, SurvPath, achieves state-of-the-art performance when evaluated against unimodal and multimodal baselines on five datasets from The Cancer Genome Atlas. Our interpretability framework identifies key multimodal prognostic factors, and, as such, can provide valuable insights into the interaction between genotype and phenotype.

Recent advancements in Spatial Transcriptomics (ST) technology have facilitated detailed gene expression analysis within tissue contexts. However, the high costs and methodological limitations of ST necessitate a more robust predictive model. In response, this paper introduces TRIPLEX, a novel deep learning framework designed to predict spatial gene expression from Whole Slide Images (WSIs). TRIPLEX uniquely harnesses multi-resolution features, capturing cellular morphology at individual spots, the local context around these spots, and the global tissue organization. By integrating these features through an effective fusion strategy, TRIPLEX achieves accurate gene expression prediction. Our comprehensive benchmark study, conducted on three public ST datasets and supplemented with Visium data from 10X Genomics, demonstrates that TRIPLEX outperforms current state-of-the-art models in Mean Squared Error (MSE), Mean Absolute Error (MAE), and Pearson Correlation Coefficient (PCC). The model's predictions align closely with ground truth gene expression profiles and tumor annotations, underscoring TRIPLEX's potential in advancing cancer diagnosis and treatment.

Teeth localization, segmentation, and labeling in 2D images have great potential in modern dentistry to enhance dental diagnostics, treatment planning, and population-based studies on oral health. However, general instance segmentation frameworks are incompetent due to 1) the subtle differences between some teeth' shapes (e.g., maxillary first premolar and second premolar), 2) the teeth's position and shape variation across subjects, and 3) the presence of abnormalities in the dentition (e.g., caries and edentulism). To address these problems, we propose a ViT-based framework named TeethSEG, which consists of stacked Multi-Scale Aggregation (MSA) blocks and an Anthropic Prior Knowledge (APK) layer. Specifically, to compose the two modules, we design 1) a unique permutation-based upscaler to ensure high efficiency while establishing clear segmentation boundaries with 2) multi-head self/cross-gating layers to emphasize particular semantics meanwhile maintaining the divergence between token embeddings. Besides, we collect 3) the first open-sourced intraoral image dataset IO150K, which comprises over 150k intraoral photos, and all photos are annotated by orthodontists using a human-machine hybrid algorithm. Experiments on IO150K demonstrate that our TeethSEG outperforms the state-of-the-art segmentation models on dental image segmentation.

The popularity of large-scale pre-training has promoted the development of medical foundation models. However, some studies have shown that although foundation models exhibit strong general feature extraction capabilities, their performance on specific tasks is still inferior to task-specific methods. In this paper, we explore a new perspective called ``Knowledge Decomposition'' to improve the performance on specific medical tasks, which deconstruct the foundation model into multiple lightweight expert models, each dedicated to a particular task, with the goal of improving specialization while concurrently mitigating resource expenditure. To accomplish the above objective, we design a novel framework named Low-Rank Knowledge Decomposition (LoRKD), which explicitly separates graidents by incorporating low-rank expert modules and the efficient knowledge separation convolution. Extensive experimental results demonstrate that the decomposed models perform well in terms of performance and transferability, even surpassing the original foundation models.

The anatomical structure detection of fetal cardiac views is crucial for diagnosing fetal congenital heart disease. In practice, there is a large domain gap between different hospitals' data, such as the variable data quality due to differences in acquisition equipment.In addition, accurate annotation information provided by obstetrician experts is always very costly or even unavailable.This study explores the unsupervised domain adaptive fetal cardiac structure detection issue.Existing unsupervised domain adaptive object detection (UDAOD) approaches mainly focus on detecting objects in natural scenes, such as Foggy Cityscapes, where the structural relationships of natural scenes are uncertain. Unlike all previous UDAOD scenarios, we first collected a $\textbf{F}$etal $\textbf{C}$ardiac $\textbf{S}$tructure dataset from three hospital centers, called $\textbf{FCS}$, and proposed a multi-matching UDA approach ($M^3$-UDA}), including histogram $\textbf{M}$atching (HM), sub-structure $\textbf{M}$atching (SM), and global-structure $\textbf{M}$atching (GM), to better transfer the topological knowledge of anatomical structure for UDA detection in medical scenarios.HM mitigates the domain gap between the source and target caused by pixel transformation. SM fuses the different angle information of the sub-organ to obtain the local topological knowledge for bridging the domain gap of the internal sub-structure.GM is designed to align the global topological knowledge of the whole organ from the source and target domain.Extensive experiments on our collected FCS and CardiacUDA, and experimental results show that $M^3$-UDA outperforms existing UDAOD studies significantly. This proposed benchmark will open new opportunities for UDAOD.

In the realm of medical 3D data, such as CT and MRI images, prevalent anisotropic resolution is characterized by high intra-slice but diminished inter-slice resolution. The lowered resolution between adjacent slices poses challenges, hindering optimal viewing experiences and impeding the development of robust downstream analysis algorithms. Various volumetric super-resolution algorithms aim to surmount these challenges, enhancing inter-slice resolution and overall 3D medical imaging quality. However, existing approaches confront inherent challenges: 1) often tailored to specific upsampling factors, lacking flexibility for diverse clinical scenarios; 2) newly generated slices frequently suffer from over-smoothing, degrading fine details, and leading to inter-slice inconsistency. In response, this study presents CycleINR, a novel enhanced Implicit Neural Representation model for 3D medical data volumetric super-resolution. Leveraging the continuity of the learned implicit function, the CycleINR model can achieve results with arbitrary up-sampling rates, eliminating the need for separate training. Additionally, we enhance the grid sampling in CycleINR with a local attention mechanism and mitigate over-smoothing by integrating cycle-consistent loss. We introduce a new metric, Slice-wise Noise Level Inconsistency (SNLI), to quantitatively assess inter-slice noise level inconsistency. The effectiveness of our approach is demonstrated through image quality evaluations on an in-house dataset and a downstream task analysis on the Medical Segmentation Decathlon liver tumor dataset.

Both limited annotation and domain shift are prevalent challenges in medical image segmentation. Traditional semi-supervised segmentation and unsupervised domain adaptation methods address one of these issues separately. However, the coexistence of limited annotation and domain shift is quite common, which motivates us to introduce a novel and challenging scenario: \textbf{Mi}xed \textbf{D}omain \textbf{S}emi-supervised medical image \textbf{S}egmentation (MiDSS). In this scenario, we handle data from multiple medical centers, with limited annotations available for a single domain and a large amount of unlabeled data from multiple domains. We found that the key to solving the problem lies in how to generate reliable pseudo labels for the unlabeled data in the presence of domain shift with labeled data. To tackle this issue, we employ Unified Copy-Paste (UCP) between images to construct intermediate domains, facilitating the knowledge transfer from the domain of labeled data to the domains of unlabeled data. To fully utilize the information within the intermediate domain, we propose a symmetric Guidance training strategy (SymGD), which additionally offers direct guidance to unlabeled data by merging pseudo labels from intermediate samples. Subsequently, we introduce a Training Process aware Random Amplitude MixUp (TP-RAM) to progressively incorporate style-transition components into intermediate samples. Compared with existing state-of-the-art approaches, our method achieves a notable 13.57\% improvement in Dice score on Prostate dataset, as demonstrated on three public datasets. Our code is available in the supplementary material.

Current vision-language pre-training (VLP) methodologies predominantly depend on paired image-text datasets, a resource that is challenging to acquire in radiology due to privacy considerations and labelling complexities. Data augmentation provides a practical solution to overcome the issue of data scarcity, however, most augmentation methods exhibit a limited focus, prioritising either image or text augmentation exclusively. Acknowledging this limitation, our objective is to devise a framework capable of concurrently augmenting medical image and text data. We design a Pairwise Augmentation (PairAug) approach that contains an Inter-patient Augmentation (InterAug) branch and an Intra-patient Augmentation (IntraAug) branch. Specifically, the InterAug branch of our approach generates radiology images using synthesised yet plausible reports derived from a Large Language Model (LLM). The generated pairs can be considered a collection of new patient cases since they are artificially created and may not exist in the original dataset.In contrast, the IntraAug branch uses newly generated reports to manipulate images. This process allows us to create new paired data for each individual with diverse medical conditions. Our extensive experiments on various downstream tasks covering medical image classification zero-shot and fine-tuning analysis demonstrate that our PairAug, concurrently expanding both image and text data, substantially outperforms image-/text-only expansion baselines and advanced medical VLP baselines.

Surgical decisions are informed by aligning rapid portable 2D intraoperative images (e.g., X-rays) to a high-fidelity 3D preoperative reference scan (e.g., CT). 2D/3D rigid registration often fails in practice: conventional optimization methods are prohibitively slow and susceptible to local minima and neural networks trained on small datasets fail on new patients or require impractical landmark supervision. We present DiffPose, a self-supervised approach that leverages patient-specific simulation and differentiable physics-based rendering to achieve accurate 2D/3D registration without relying on manually labeled data. Preoperatively, a CNN is trained to regress the pose of a randomly oriented synthetic X-ray rendered from the preoperative CT. The CNN then initializes rapid intraoperative test-time optimization that uses the differentiable X-ray renderer to refine the solution. Our work further proposes several geometrically principled methods for sampling camera poses from $\mathbf{SE}(3)$, for sparse differentiable rendering, and for driving registration in the tangent space $\mathfrak{se}(3)$ with geodesic and multiscale locality-sensitive losses. As a result, DiffPose achieves sub-millimeter accuracy across surgical datasets at intraoperative speeds, improving upon existing unsupervised methods by an order of magnitude and even outperforming supervised baselines. Code is provided.

Nuclear instance segmentation has played a critical role in pathology image analysis. The main challenges arise from the difficulty in accurately segmenting densely overlapping instances and the high cost of precise mask-level annotations. Existing fully-supervised nuclear instance segmentation methods, such as boundary-based methods, struggle to capture differences between overlapping instances and thus fail in densely distributed blurry regions. They also face challenges transitioning to point supervision, where annotations are simple and effective. Inspired by natural mudslides, we propose a universal method called Mudslide that uses simple representations to characterize differences between different instances and can easily be extended from fully-supervised to point-supervised. Concretely, we introduce a collapse field and leverage it to construct a force map and initial boundary, enabling a distinctive representation for each instance. Each pixel is assigned a collapse force, with distinct directions between adjacent instances. Starting from the initial boundary, Mudslide executes a pixel-by-pixel collapse along various force directions. Pixels that collapse into the same region are considered as one instance, concurrently accounting for both inter-instance distinctions and intra-instance coherence. Experiments on public datasets show superior performance in both fully-supervised and point-supervised tasks. The code is available at https://github.com/CVPR2024/mudslide.

Deformable image registration is a fundamental step for medical image analysis. Recently, transformers have been used for registration and outperformed Convolutional Neural Networks (CNNs). Transformers can capture long-range dependence among image features, which have been shown beneficial for registration. However, due to the high computation/memory loads of self-attention, transformers are typically used at downsampled feature resolutions and cannot capture fine-grained long-range dependence at the full image resolution. This limits deformable registration as it necessitates precise dense correspondence between each image pixel. Multi-layer Perceptrons (MLPs) without self-attention are efficient in computation/memory usage, enabling the feasibility of capturing fine-grained long-range dependence at full resolution. Nevertheless, MLPs have not been extensively explored for image registration and are lacking the consideration of inductive bias crucial for medical registration tasks. In this study, we propose the first correlation-aware MLP-based registration network (CorrMLP) for deformable medical image registration. Our CorrMLP introduces a correlation-aware multi-window MLP block in a novel coarse-to-fine registration architecture, which captures fine-grained multi-range dependence to perform correlation-aware coarse-to-fine registration. Extensive experiments with seven public medical datasets show that our CorrMLP outperforms state-of-the-art deformable registration methods.

This paper addresses complex challenges in histopathological image analysis through three key contributions. Firstly, it introduces a fast patch selection method, FPS, for whole-slide image (WSI) analysis, significantly reducing computational cost while maintaining accuracy. Secondly, it presents PathDino, a lightweight histopathology feature extractor with a minimal configuration of five Transformer blocks and only 9 million parameters, markedly fewer than alternatives. Thirdly, it introduces a rotation-agnostic representation learning paradigm using self-supervised learning, effectively mitigating overfitting. We also show that our compact model outperforms existing state-of-the-art histopathology-specific vision transformers on 12 diverse datasets, including both internal datasets spanning four sites (breast, liver, skin, and colorectal) and seven public datasets (PANDA, CAMELYON16, BRACS, DigestPath, Kather, PanNuke, and WSSS4LUAD). Notably, even with a training dataset of 6 million histopathology patches from The Cancer Genome Atlas (TCGA), our approach demonstrates an average 8.5% improvement in patch-level majority vote performance. These contributions provide a robust framework for enhancing image analysis in digital pathology, rigorously validated through extensive evaluation.

The recent advance of deep learning technology brings the possibility of assisting the pathologist to predict the patients’ survival from whole-slide pathological images(WSIs). However, most of the prevalent methods only worked on the sampled patches in specifically or randomly selected tumor areas of WSIs, which has very limited capability to capture the complex interactions between tumorand its surrounding micro-environment components. As a matter of fact, tumor is supported and nurtured in the heterogeneous tumor micro-environment(TME), and the detailed analysis of TME and their correlation with tumors are important to in-depth analyze the mechanism of cancer development. In this paper, we considered the spatial interactions among tumor and its two major TME components(i.e., lymphocytes and stromal fibrosis) and proposed a Tumor Micro-environment Interactions Guided Graph Learnng (TMEGL) algorithm for the prognosis prediction of human cancers. Specifically, we firstly selected different types of patches as nodes to build graph for each WSI. Then, a novel TME neighborhood organization guided graph embedding algorithm was proposed to learn node representations that can preserve their topological structure information. Finally, a Gated Graph Attention Network is applied to capture survival-associated intersections among tumor and different TME components for clinical outcome prediction. We tested TMEGL on three cancer cohorts derived from The Cancer Genome Atlas (TCGA), and the experimental results indicated that TMEGL not only outperforms the existing WSI-based survival analysis models, but also has good explainable ability for survival prediction.

The scarcity of annotated data has sparked significant interest in unsupervised pre-training methods that leverage medical reports as auxiliary signals for medical visual representation learning. However, existing research overlooks the multi-granularity nature of medical visual representation and lacks suitable contrastive learning techniques to improve the models' generalizability across different granularities, leading to the underutilization of image-text information. To address this, we propose MLIP, a novel framework leveraging domain-specific medical knowledge as guiding signals to integrate language information into the visual domain through image-text contrastive learning. Our model includes global contrastive learning with our designed divergence encoder, local token-knowledge-patch alignment contrastive learning, and knowledge-guided category-level contrastive learning with expert knowledge. Experimental evaluations reveal the efficacy of our model in enhancing transfer performance for tasks such as image classification, object detection, and semantic segmentation. Notably, MLIP surpasses state-of-the-art methods even with limited annotated data, highlighting the potential of multimodal pre-training in advancing medical representation learning.

In recent years, automated Gallbladder Cancer (GBC) detection has gained the attention of researchers. Current state-of-the-art (SOTA) methodologies relying on ultrasound sonography (US) images exhibit limited generalization, emphasizing the need for transformative approaches. We observe that individual US frames may lack sufficient information to capture disease manifestation. This study advocates for a paradigm shift towards video-based GBC detection, leveraging the inherent advantages of spatio-temporal representations. Employing the Masked Autoencoder (MAE) for representation learning, we address shortcomings in conventional image-based methods. We propose a novel design called FocusMAE to systematically bias the selection of masking tokens from high-information regions, fostering a more refined representation of malignancy. Additionally, we contribute the most extensive US video dataset for GBC detection. We also note that, this is the first study on US video-based GBC detection. We validate the proposed methods on the curated dataset, and report a new state-of-the-art (SOTA) accuracy of 96.4% for the GBC detection problem, against an accuracy of 84% by current SOTA - GBCNet, and RadFormer. We further demonstrate the generality of the proposed FocusMAE on a public CT-based Covid detection dataset, reporting an improvement in accuracy by 2.2% over current baselines. The source code and pre-trained models are available.

One-shot medical image segmentation (MIS) aims to cope with the expensive, time-consuming, and inherent human bias annotations. One prevalent method to address one-shot MIS is joint registration and segmentation (JRS) with a shared encoder, which mainly explores the voxel-wise correspondence between the labeled data and unlabeled data for better segmentation. However, this method omits underlying connections between task-specific decoders for segmentation and registration, leading to unstable training. In this paper, we propose a novel Bi-level Learning of Task-Specific Decoders for one-shot MIS, employing a pretrained fixed shared encoder that is proved to be morequickly adapted to brand-new datasets than existing JRS without fixed shared encoder paradigm. To be more specific, we introduce a bi-level optimization training strategy considering registration as a major objective and segmentation as a learnable constraint by leveraging inter-task coupling dependencies. Furthermore, we design an appearance conformity constraint strategy that learns the backward transformations generating the fake labeled data used to perform data augmentation instead of the labeled image, to avoid performance degradation caused by inconsistent styles between unlabeled data and labeled data in previous methods. Extensive experiments on the brain MRI task across ABIDE, ADNI, and PPMI datasets demonstrate that the proposed Bi-JROS outperforms state-of-the-art one-shotMIS methods for both segmentation and registration tasks. The code will be available at https://github.com/Coradlut/Bi-JROS.

Understanding the anatomy of renal pathology is crucial for advancing disease diagnostics, treatment evaluation, and clinical research. The complex kidney system comprises various components across multiple levels, including regions (cortex, medulla), functional units (glomeruli, tubules), and cells (podocytes, mesangial cells in glomerulus). Prior studies have predominantly overlooked the intricate spatial interrelations among objects from clinical knowledge. In this research, we introduce a novel universal proposition learning approach, called panoramic renal pathology segmentation (PrPSeg), designed to segment comprehensively panoramic structures within kidney by integrating extensive knowledge of kidney anatomy.In this paper, we propose (1) the design of a comprehensive universal proposition matrix for renal pathology, facilitating the incorporation of classification and spatial relationships into the segmentation process; (2) a token-based dynamic head single network architecture, with the improvement of the partial label image segmentation and capability for future data enlargement; and (3) an anatomy loss function, quantifying the inter-object relationships across the kidney.

A versatile medical image segmentation model applicable to imaging data collected with diverse equipment and protocols can facilitate model deployment and maintenance. However, building such a model typically requires a large, diverse, and fully annotated dataset, which is rarely available due to the labor-intensive and costly data curation. In this study, we develop a cost-efficient method by harnessing readily available data with partially or even sparsely annotated segmentation labels. We devise strategies for model self-disambiguation, prior knowledge incorporation, and imbalance mitigation to address challenges associated with inconsistently labeled data from various sources, including label ambiguity and imbalances across modalities, datasets, and segmentation labels. Experimental results on a multi-modal dataset compiled from eight different sources for abdominal organ segmentation have demonstrated our method's effectiveness and superior performance over alternative state-of-the-art methods, highlighting its potential for optimizing the use of existing annotated data and reducing the annotation efforts for new data to further enhance model capability.

Featurizing microscopy images for use in biological research remains a significant challenge, especially for large-scale experiments spanning millions of images. This work explores the scaling properties of weakly supervised classifiers and self-supervised masked autoencoders (MAEs) when training with increasingly larger model backbones and microscopy datasets. Our results show that ViT-based MAEs outperform weakly supervised classifiers on a variety of tasks, achieving as much as a 11.5% relative improvement when recalling known biological relationships curated from public databases. Additionally, we develop a new channel-agnostic MAE architecture (CA-MAE) that allows for inputting images of different numbers and orders of channels at inference time. We demonstrate that CA-MAEs effectively generalize by inferring and evaluating on a microscopy image dataset (JUMP-CP) generated under different experimental conditions with a different channel structure than our pretraining data (RPI-93M). Our findings motivate continued research into scaling self-supervised learning on microscopy data in order to create powerful foundation models of cellular biology that have the potential to catalyze advancements in drug discovery and beyond.

An efficient and effective decoding mechanism is crucial in medical image segmentation, especially in scenarios with limited computational resources. However, these decoding mechanisms usually come with high computational costs. To address this concern, we introduce EMCAD, a new efficient multi-scale convolutional attention decoder, designed to optimize both performance and computational efficiency. EMCAD leverages a unique multi-scale depth-wise convolution block, significantly enhancing feature maps through multi-scale convolutions. EMCAD also employs channel, spatial, and grouped (large-kernel) gated attention mechanisms, which are highly effective at capturing intricate spatial relationships while focusing on salient regions. By employing group and depth-wise convolution, EMCAD is very efficient and scales well (e.g., only 1.91M parameters and 0.381G FLOPs are needed when using a standard encoder). Our rigorous evaluations across 12 datasets that belong to six medical image segmentation tasks reveal that EMCAD achieves state-of-the-art (SOTA) performance with 79.4% and 80.3% reduction in #Params and #FLOPs, respectively. Moreover, EMCAD’s adaptability to different encoders and versatility across segmentation tasks further establish EMCAD as a promising tool, advancing the field towards more efficient and accurate medical image analysis. Our implementation is available at https://github.com/SLDGroup/EMCAD.

Among the numerous efforts towards digitally recovering the physical world, Neural Radiance Fields (NeRFs) have proved effective in most cases. However, underwater scene introduces unique challenges due to the absorbing water medium, the local change in lighting and the dynamic contents in the scene. We aim at developing a neural underwater scene representation for these challenges, modeling the complex process of attenuation, unstable in-scattering and moving objects during light transport. The proposed method can reconstruct the scenes from both established datasets and in-the-wild videos with outstanding fidelity.

Recent years have seen immense progress in 3D computer vision and computer graphics, with emerging tools that can virtualize real-world 3D environments for numerous Mixed Reality (XR) applications. However, alongside immersive visual experiences, immersive auditory experiences are equally vital to our holistic perception of an environment. In this paper, we aim to reconstruct the spatial acoustic characteristics of an arbitrary environment given only a sparse set of (roughly 12) room impulse response (RIR) recordings and a planar reconstruction of the scene, a setup that is easily achievable by ordinary users. To this end, we introduce DiffRIR, a differentiable RIR rendering framework with interpretable parametric models of salient acoustic features of the scene, including sound source directivity and surface reflectivity. This allows us to synthesize novel auditory experiences through the space with any source audio. To evaluate our method, we collect a dataset of RIR recordings and music in four diverse, real environments. We show that our model outperforms state-of-the-art baselines on rendering monaural and binaural RIRs and music at unseen locations, and learns physically interpretable parameters characterizing acoustic properties of the sound source and surfaces in the scene.

This paper introduces a versatile multi-view inverse rendering framework with near- and far-field light sources. Tackling the fundamental challenge of inherent ambiguity in inverse rendering, our framework adopts a lightweight yet inclusive lighting model for different near- and far-field lights, thus is able to make use of input images under varied lighting conditions available during capture.It leverages observations under each lighting to disentangle the intrinsic geometry and material from the external lighting, using both neural radiance field rendering and physically-based surface rendering on the 3D implicit fields.After training, the reconstructed scene is extracted to a textured triangle mesh for seamless integration into industrial rendering software for various applications.Quantitatively and qualitatively tested on synthetic and real-world scenes, our method shows superiority to state-of-the-art multi-view inverse rendering methods in both speed and quality.

Photometric stereo is a well-established technique to estimate the surface normal of an object. However, the requirement of capturing multiple high dynamic range images under different illumination conditions limits the speed and real-time applications. This paper introduces EventPS, a novel approach to real-time photometric stereo using an event camera. Capitalizing on the exceptional temporal resolution, dynamic range, and low bandwidth characteristics of event cameras, EventPS estimates surface normal only from the radiance changes, significantly enhancing data efficiency. EventPS seamlessly integrates with both optimization-based and deep-learning-based photometric stereo techniques to offer a robust solution for non-Lambertian surfaces. Extensive experiments validate the effectiveness and efficiency of EventPS compared to frame-based counterparts. Our algorithm runs at over 30 fps in real-world scenarios, unleashing the potential of EventPS in time-sensitive and high-speed downstream applications.

Photometric stereo faces challenges from non-Lambertian reflectance in real-world scenarios. Systematically measuring the reliability of photometric stereo methods in handling such complex reflectance necessitates a real-world dataset with quantitatively controlled reflectances. This paper introduces DiLiGenRT, the first real-world dataset for evaluating photometric stereo methods under quantified reflectances by manufacturing 54 hemispheres with varying degrees of two reflectance properties: Roughness and Translucency. Unlike qualitative and semantic labels, such as diffuse and specular, that have been used in previous datasets, our quantified dataset allows comprehensive and systematic benchmark evaluations. In addition, it facilitates selecting best-fit photometric stereo methods based on the quantitative reflectance properties. Our dataset and benchmark results are available at https://photometricstereo.github.io/diligentrt.html.

We present NeRSP, a Neural 3D reconstruction technique for Reflective surfaces with Sparse Polarized images. Reflective surface reconstruction is extremely challenging as specular reflections are view-dependent and thus violate the multiview consistency for multiview stereo. On the other hand, sparse image inputs, as a practical capture setting, commonly cause incomplete or distorted results due to the lack of correspondence matching. This paper jointly handles the challenges from sparse inputs and reflective surfaces by leveraging polarized images. We derive photometric and geometric cues from the polarimetric image formation model and multiview azimuth consistency, which jointly optimize the surface geometry modeled via implicit neural representation. Based on the experiments on our synthetic and real datasets, we achieve the state-of-the-art surface reconstruction results with only 6 views as input.

Separating the direct and global components of a scene aids in shape recovery and basic material understanding. Conventional methods capture multiple frames under high frequency illumination patterns or shadows, requiring the scene to keep stationary during the image acquisition process. Single-frame methods simplify the capture procedure but yield lower-quality separation results. In this paper, we leverage the event camera to facilitate the separation of direct and global components, enabling video-rate separation of high quality. In detail, we adopt an event camera to record rapid illumination changes caused by the shadow of a line occluder sweeping over the scene, and reconstruct the coarse separation results through event accumulation. We then design a network to resolve the noise in the coarse separation results and restore color information. A real-world dataset is collected using a hybrid camera system for network training and evaluation. Experimental results show superior performance over state-of-the-art methods.

Photometric stereo leverages variations in illumination conditions to reconstruct surface normals.Display photometric stereo, which employs a conventional monitor as an illumination source, has the potential to overcome limitations often encountered in bulky and difficult-to-use conventional setups. In this paper, we present differentiable display photometric stereo (DDPS), addressing an often overlooked challenge in display photometric stereo: the design of display patterns. Departing from using heuristic display patterns, DDPS learns the display patterns that yield accurate normal reconstruction for a target system in an end-to-end manner.To this end, we propose a differentiable framework that couples basis-illumination image formation with analytic photometric-stereo reconstruction. The differentiable framework facilitates the effective learning of display patterns via auto-differentiation.Also, for training supervision, we propose to use 3D printing for creating a real-world training dataset, enabling accurate reconstruction on the target real-world setup.Finally, we exploit that conventional LCD monitors emit polarized light, which allows for the optical separation of diffuse and specular reflections when combined with a polarization camera, leading to accurate normal reconstruction.Extensive evaluation of DDPS shows improved normal-reconstruction accuracy compared to heuristic patterns and demonstrates compelling properties such as robustness to pattern initialization, calibration errors, and simplifications in image formation and reconstruction.

We propose a new method for cloth digitalization. Aiming for accurate cloth digitalization, we deviate from existing methods which learn from data captured under relatively casual settings. We propose to learn from data captured in strictly tested measuring protocols, and find plausible physical parameters of the cloths. However, such data is currently absent, so we first propose a new dataset with accurate cloth measurements. Further, the data size is considerably smaller than the ones used in current deep learning, due to the nature of the data capture process. To be able to learn from small data, we propose a new Bayesian differentiable cloth model to estimate and capture the complex material heterogeneity of real cloths. It can provide highly accurate digitalization from very limited data samples. Through exhaustive evaluations and comparisons, we show our method is accurate in cloth digitalization, efficient in learning from limited data samples, and general in capturing material variations.

Monocular depth estimation has experienced significant progress on terrestrial images in recent years thanks to deep learning advancements. But it remains inadequate for underwater scenes primarily due to data scarcity. Given the inherent challenges of light attenuation and backscatter in water, acquiring clear underwater images or precise depth is notably difficult and costly. To mitigate this issue, learning-based approaches often rely on synthetic data or turn to self- or unsupervised manners. Nonetheless, their performance is often hindered by domain gap and looser constraints. In this paper, we propose a novel pipeline for generating photorealistic underwater images using accurate terrestrial depth. This approach facilitates the supervised training of models for underwater depth estimation, effectively reducing the performance disparity between terrestrial and underwater environments. Contrary to previous synthetic datasets that merely apply style transfer to terrestrial images without scene content change, our approach uniquely creates vivid non-existent underwater scenes by leveraging terrestrial depth data through the innovative Stable Diffusion model. Specifically, we introduce a specialized Depth2Underwater ControlNet, trained on prepared {Underwater, Depth, Text} data triplets, for this generation task. Our newly developed dataset, Atlantis, enables terrestrial depth estimation models to achieve considerable improvements on unseen underwater scenes, surpassing their terrestrial pretrained counterparts both quantitatively and qualitatively. Moreover, we further show its practical utility by applying the improved depth in underwater image enhancement, and its smaller domain gap from the LLVM perspective. Code and dataset are publicly available at https://github.com/zkawfanx/Atlantis.

Neural approaches have shown a significant progress on camera-based reconstruction. But they require either a fairly dense sampling of the viewing sphere, or pre-training on an existing dataset, thereby limiting their generalizability. In contrast, photometric stereo (PS) approaches have shown great potential for achieving high-quality reconstruction under sparse viewpoints. Yet, they are impractical because they typically require tedious laboratory conditions, are restricted to dark rooms, and often multi-staged, making them subject to accumulated errors. To address these shortcomings, we propose an end-to-end uncalibrated multi-view PS framework for reconstructing high-resolution shapes acquired from sparse viewpoints in a real-world environment. We relax the dark room assumption, and allow a combination of static ambient lighting and dynamic near LED lighting, thereby enabling easy data capture outside the lab. Experimental validation confirms that it outperforms existing baseline approaches in the regime of sparse viewpoints by a large margin. This allows to bring high-accuracy 3D reconstruction from the dark room to the real world, while maintaining a reasonable data capture complexity.

Reflectance bounds the frequency spectrum of illumination in the object appearance. In this paper, we introduce the first stochastic inverse rendering method, which recovers the attenuated frequency spectrum of an illumination jointly with the reflectance of an object of known geometry from a single image. Our key idea is to solve this blind inverse problem in the reflectance map, an appearance representation invariant to the underlying geometry, by learning to reverse the image formation with a novel diffusion model which we refer to as the Diffusion Reflectance Map Network (DRMNet). Given an observed reflectance map converted and completed from the single input image, DRMNet generates a reflectance map corresponding to a perfect mirror sphere while jointly estimating the reflectance. The forward process can be understood as gradually filtering a natural illumination with lower and lower frequency reflectance and additive Gaussian noise. DRMNet learns to invert this process with two subnetworks, IllNet and RefNet, which work in concert towards this joint estimation. The network is trained on an extensive synthetic dataset and is demonstrated to generalize to real images, showing state-of-the-art accuracy on established datasets.

A Manhattan world lying along cuboid buildings is useful for camera angle estimation. However, accurate and robust angle estimation from fisheye images in the Manhattan world has remained an open challenge because general scene images tend to lack constraints such as lines, arcs, and vanishing points. To achieve higher accuracy and robustness, we propose a learning-based calibration method that uses heatmap regression, which is similar to pose estimation using keypoints, to detect the directions of labeled image coordinates. Simultaneously, our two estimators recover the rotation and remove fisheye distortion by remapping from a general scene image. Without considering vanishing-point constraints, we find that additional points for learning-based methods can be defined. To compensate for the lack of vanishing points in images, we introduce auxiliary diagonal points that have the optimal 3D arrangement of spatial uniformity. Extensive experiments demonstrated that our method outperforms conventional methods on large-scale datasets and with off-the-shelf cameras.

3D reconstruction of dynamic scenes is a long-standing problem in computer graphics and increasingly difficult the less information is available. Shape-from-Template (SfT) methods aim to reconstruct a template-based geometry from RGB images or video sequences, often leveraging just a single monocular camera without depth information, such as regular smartphone recordings. Unfortunately, existing reconstruction methods are either unphysical and noisy or slow in optimization. To solve this problem, we propose a novel SfT reconstruction algorithm for cloth using a pre-trained neural surrogate model that is fast to evaluate, stable, and produces smooth reconstructions due to a regularizing physics simulation. Differentiable rendering of the simulated mesh enables pixel-wise comparisons between the reconstruction and a target video sequence that can be used for a gradient-based optimization procedure to extract not only shape information but also physical parameters such as stretching, shearing, or bending stiffness of the cloth. This allows to retain a precise, stable, and smooth reconstructed geometry while reducing the runtime by a factor of 400–500 compared to $\phi$-SfT, a state-of-the-art physics-based SfT approach.

Natural Light Uncalibrated Photometric Stereo (NaUPS) relieves the strict environment and light assumptions in classical Uncalibrated Photometric Stereo (UPS) methods. However, due to the intrinsic ill-posedness and high-dimensional ambiguities, addressing NaUPS is still an open question. Existing works impose strong assumptions on the environment lights and objects’ material, restricting the effectiveness in more general scenarios. Alternatively, some methods leverage supervised learning with intricate models while lacking interpretability, resulting in a biased estimation. In this work, we proposed Spin Light Uncalibrated Photometric Stereo (Spin-UP), an unsupervised method to tackle NaUPS in various environment lights and objects. The proposed method uses a novel setup that captures the object’s images on a spinning platform, which mitigates NaUPS’s ill-posedness by reducing unknowns and provides reliable priors to alleviate NaUPS’s ambiguities. Leveraging neural inverse rendering and the proposed training strategies, Spin-UP recovers surface normals, environment light, and isotropic reflectance under complex natural light. Experiments have shown that Spin-UP outperforms other supervised / unsupervised NaUPS methods and achieves state-of-the-art performance on synthetic and real-world datasets. Codes and data are available at https://github.com/LMozart/CVPR2024-SpinUP.

Many surface reconstruction methods incorporate normal integration, which is a process to obtain a depth map from surface gradients. In this process, the input may represent a surface with discontinuities, e.g., due to self-occlusion. To reconstruct an accurate depth map from the input normal map, hidden surface gradients occurring from the jumps must be handled. To model these jumps correctly, we design a novel discretization for the domain of normal integration. Our key idea is to introduce auxiliary edges, which bridge between piecewise-smooth planes in the domain so that the magnitude of hidden jumps can be explicitly expressed on finite elements. Using the auxiliary edges, we design a novel algorithm to optimize the discontinuity and the depth map from the input normal map. Our method optimizes discontinuities by using a combination of iterative re-weighted least squares and iterative filtering of the jump magnitudes on auxiliary edges to provide strong sparsity regularization. Compared to previous discontinuity-preserving normal integration methods, which model the magnitude of jumps only implicitly, our method reconstructs subtle discontinuities accurately thanks to our explicit representation allowing for strong sparsity regularization.

We present a novel theory that establishes the relationship between light transport in visible and thermal infrared, and heat transport in solids. We show that heat generated due to light absorption can be estimated by modeling heat transport using a thermal camera. For situations where heat conduction is negligible, we analytically solve the heat transport equation to derive a simple expression relating the change in thermal image intensity to the absorbed light intensity and heat capacity of the material. Next, we prove that intrinsic image decomposition for Lambertian scenes becomes a well-posed problem if one has access to the absorbed light. Our theory generalizes to arbitrary shapes and unstructured illumination. Our theory is based on applying energy conservation principle at each pixel independently. We validate our theory using real-world experiments on diffuse objects made of different materials that exhibit both direct and global components (inter-reflections) of light transport under unknown complex lighting.

In this work, we propose IDGuard, a novel proactive defense method from the perspective of developers, to protect Persons-of-Interest (POI) such as national leaders from face editing abuse. We build a bridge between identities and model behavior, safeguarding POI identities rather than merely shielding certain face images. Given a face editing model, IDGuard enables it to reject editing any image containing POI identities while retaining its editing functionality for regular use. Specifically, we insert an ID Normalization Layer into the original face editing model and introduce an ID Extractor to extract the identities of input images. To differentiate the editing behavior between POI and nonPOI, we use a transformer-based ID Encoder to encode extracted POI identities as parameters of the ID Normalization Layer. Our method supports the simultaneous protection of multiple POI and allows for the addition of new POI in the inference stage, without the need for retraining. Extensive experiments show that our method achieves 100% protection accuracy on POI images even if they are neither included in the training set nor subject to any preprocessing. Notably, our method exhibits excellent robustness against image and model attacks and maintains 100% protection performance when generalized to various face editing models, further demonstrating its practicality.

The training of contemporary deep learning models heavily relies on publicly available data, posing a risk of unauthorized access to online data and raising concerns about data privacy. Current approaches to creating unlearnable data involve incorporating small, specially designed noises, but these methods strictly limit data usability, overlooking its potential usage in authorized scenarios.In this paper, we extend the concept of unlearnable data to conditional data learnability and introduce \textbf{U}n\textbf{G}eneralizable \textbf{E}xamples (UGEs). UGEs exhibit learnability for authorized users while maintaining unlearnability for potential hackers. The protector defines the authorized network and optimizes UGEs to match the gradients of the original data and its ungeneralizable version, ensuring learnability. To prevent unauthorized learning, UGEs are trained by maximizing a designated distance loss in a common feature space. Additionally, to further safeguard the authorized side from potential attacks, we introduce additional undistillation optimization.Experimental results on multiple datasets and various networks demonstrate that the proposed UGEs framework preserves data usability while reducing training performance on hacker networks, even under different types of attacks.

The proliferation of large-scale AI models trained on extensive datasets has revolutionized machine learning. With these models taking on increasingly central roles in various applications, the need to understand their behavior and enhance interpretability has become paramount.To investigate the impact of changes in training data on a pre-trained model, a common approach is leave-one-out retraining. This entails systematically altering the training dataset by removing specific samples to observe resulting changes within the model. However, retraining the model for each altered dataset presents a significant computational challenge, given the need to perform this operation for every dataset variation.In this paper, we introduce an efficient framework for assessing data impact, comprising offline training and online evaluation stages. During the offline training phase, we approximate the influence of training data on the target model through a distilled synset, formulated as a reversed gradient matching problem. For online evaluation, we expedite the leave-one-out process using the synset, which is then utilized to compute the attribution matrix based on the evaluation objective.Experimental evaluations, including training data attribution and assessments of data quality, demonstrate that our proposed method achieves comparable model behavior evaluation while significantly speeding up the process compared to the direct retraining method.

In the era of AI-generated content (AIGC), malicious tampering poses imminent threats to copyright integrity and information security. Current deep image watermarking, while widely accepted for safeguarding visual content, can only protect copyright and ensure traceability. They fall short in localizing increasingly realistic image tampering, potentially leading to trust crises, privacy violations, and legal disputes. To solve this challenge, we propose an innovative proactive forensics framework EditGuard, to unify copyright protection and tamper-agnostic localization, especially for AIGC-based editing methods. It can offer a meticulous embedding of imperceptible watermarks and precise decoding of tampered areas and copyright information. Leveraging our observed fragility and locality of image-into-image steganography, the realization of EditGuard can be converted into a united image-bit steganography issue, thus completely decoupling the training process from the tampering types. Extensive experiments verify that our EditGuard balances the tamper localization accuracy, copyright recovery precision, and generalizability to various AIGC-based tampering methods, especially for image forgery that is difficult for the naked eye to detect.

While vision-language models (VLMs) have achieved remarkable performance improvements recently, there is growing evidence that these models also posses harmful biases with respect to social attributes such as gender and race. Prior studies have primarily focused on probing such bias attributes individually while ignoring biases associated with intersections between social attributes. This could be due to the difficulty of collecting an exhaustive set of image-text pairs for various combinations of social attributes. To address this challenge, we employ text-to-image diffusion models to produce counterfactual examples for probing intersectional social biases at scale. Our approach utilizes Stable Diffusion with cross attention control to produce sets of counterfactual image-text pairs that are highly similar in their depiction of a subject (e.g., a given occupation) while differing only in their depiction of intersectional social attributes (e.g., race & gender). Through our over-generate-then-filter methodology, we produce SocialCounterfactuals, a high-quality dataset containing 171k image-text pairs for probing intersectional biases related to gender, race, and physical characteristics. We conduct extensive experiments to demonstrate the usefulness of our generated dataset for probing and mitigating intersectional social biases in state-of-the-art VLMs.

Personalized Federated Learning (PFL) is primarily designed to provide customized models for each client to better fit the non-iid distributed client data, which is a inherent challenge in Federated Learning.However, current PFL methods suffer from inconsistencies in both intra-client and inter-client levels: 1) The intra-client inconsistency stems from the asynchronous update strategy for personalized and shared parameters in PFL. Client updates their shared parameters to communicate and learn from other clients, while keeping personalized parts unchanged, leading to poor coordination between these two components.2) The Inter-client inconsistency in PFL arises from “stragglers" - inactive clients that communicate and train with the server less frequently. This results in their under-trained personalized models and impedes the collaborative training stage for other clients.In this paper, we present a novel PFL framework named \textbf{FedAS}, which uses \textbf{Fed}erated Parameter-\textbf{A}lignment and Client-\textbf{S}ynchronization to overcome above challenges. Initially, we enhance the localization of global parameters by infusing them with local insights, thereby increasing their relevance and consistency with local information.To achieve this, we make the shared parts learn from previous model, and thereby reduce the impact from parameter inconsistency.Furthermore, we design a robust aggregation method to mitigate the impact of stragglers by preventing the incorporation of their under-trained knowledge into global aggregated model.Experimental results on Cifar10 and Cifar100 datasets validate the effectiveness of our FedAS in achieving better personalization performance and robustness against data heterogeneity.

Existing text-to-image generative models reflect or even amplify societal biases ingrained in their training data. This is especially concerning for human image generation where models are biased against certain demographic groups. Existing attempts to rectify this issue are hindered by the inherent limitations of the pre-trained models and fail to substantially improve demographic diversity. In this work, we introduce Fair Retrieval Augmented Generation (FairRAG), a novel framework that conditions pre-trained generative models on reference images retrieved from an external image database to improve fairness in human generation. FairRAG enables conditioning through a lightweight linear module that projects reference images into the textual space. To enhance fairness, FairRAG applies simple-yet-effective debiasing strategies, providing images from diverse demographic groups during the generative process. Extensive experiments demonstrate that FairRAG outperforms existing methods in terms of demographic diversity, image-text alignment and image fidelity while incurring minimal computational overhead during inference.

Diffusion-based models have gained significant popularity for text-to-image generation due to their exceptional image-generation capabilities. A risk with these models is the potential generation of inappropriate content, such as biased or harmful images. However, the underlying reasons for generating such undesired content from the perspective of the diffusion model's internal representation remain unclear. Previous work interprets vectors in an interpretable latent space of diffusion models as semantic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation. Extensive experiments have been conducted to verify the effectiveness of our mitigation approach, namely, for fair generation, safe generation, and responsible text-enhancing generation.

Group robustness strategies aim to mitigate learned biases in deep learning models that arise from spurious correlations present in their training datasets. However, most existing methods rely on the access to the label distribution of the groups, which is time-consuming and expensive to obtain. As a result, unsupervised group robustness strategies are sought. Based on the insight that a trained model's classification strategies can be inferred accurately based on explainability heatmaps, we introduce ExMap, an unsupervised two stage mechanism designed to enhance group robustness in traditional classifiers. ExMap utilizes a clustering module to infer pseudo-labels based on a model's explainability heatmaps, which are then used during training in lieu of actual labels. Our empirical studies validate the efficacy of ExMap - We demonstrate that it bridges the performance gap with its supervised counterparts and outperforms existing partially supervised and unsupervised methods. Additionally, ExMap can be seamlessly integrated with existing group robustness learning strategies. Finally, we demonstrate its potential in tackling the emerging issue of multiple shortcut mitigation

Federated Learning (FL) enables collaborative model training while preserving the privacy of raw data. A challenge in this framework is the fair and efficient valuation of data, which is crucial for incentivizing clients to contribute high-quality data in the FL task. In scenarios involving numerous data clients within FL, it is often the case that only a subset of clients and datasets are pertinent to a specific learning task, while others might have either a negative or negligible impact on the model training process. This paper introduces a novel privacy-preserving method for evaluating client contributions and selecting relevant datasets without a pre-specified training algorithm in an FL task. Our proposed approach, $\texttt{FedBary}$, utilizes Wasserstein distance within the federated context, offering a new solution for data valuation in the FL framework. This method ensures transparent data valuation and efficient computation of the Wasserstein barycenter and reduces the dependence on validation datasets. Through extensive empirical experiments and theoretical analysies, we demonstrate the potential of this data valuation method as a promising avenue for FL research.

When building classification systems with demographic fairness considerations, there are two objectives to satisfy: 1) maximizing utility for the specific task and 2) ensuring fairness w.r.t. a known demographic attribute. These objectives often compete, so optimizing both can lead to a trade-off between utility and fairness. While existing works acknowledge the trade-offs and study their limits, two questions remain unanswered: 1) What are the optimal trade-offs between utility and fairness? and 2) How can we numerically quantify these trade-offs from data for a desired prediction task and demographic attribute of interest? This paper addresses these questions. We introduce two utility-fairness trade-offs: the Data-Space and Label-Space Trade-off. The trade-offs reveal three regions within the utility-fairness plane, delineating what is fully and partially possible and impossible. We propose U-FaTE, a method to numerically quantify the trade-offs for a given prediction task and group fairness definition from data samples. Based on the trade-offs, we introduce a new scheme for evaluating representations. An extensive evaluation of fair representation learning methods and representations from over 1000 pre-trained models revealed that most current approaches are far from the estimated fairness-utility trade-offs across multiple datasets and prediction tasks.

Despite the success of diffusion-based customization methods on visual content creation, increasing concerns have been raised about such techniques from both privacy and political perspectives. To tackle this issue, several anti-customization methods have been proposed in very recent months, predominantly grounded in adversarial attacks. Unfortunately, most of these methods adopt straightforward designs, such as end-to-end optimization with a focus on adversarially maximizing the original training loss, thereby neglecting nuanced internal properties intrinsic to the diffusion model, and even leading to ineffective optimization in some diffusion time steps. In this paper, we strive to bridge this gap by undertaking a comprehensive exploration of these inherent properties, to boost the performance of current anti-customization approaches. Two aspects of properties are investigated: 1) We examine the relationship between time step selection and the model's perception in the frequency domain of images and find that lower time steps can give much more contributions to adversarial noises. This inspires us to propose an adaptive greedy search for optimal time steps that seamlessly integrates with existing anti-customization methods. 2) We scrutinize the roles of features at different layers during denoising and devise a sophisticated feature-based optimization framework for anti-customization. Experiments on facial benchmarks demonstrate that our approach significantly increases identity disruption, thereby protecting user privacy and copyright. Our code is available at: https://github.com/somuchtome/SimAC.

Adversarial attacks aim to perturb images such that a predictor outputs incorrect results. Due to the limited research in structured attacks, imposing consistency checks on natural multi-object scenes is a practical defense against conventional adversarial attacks. More desired attacks should be able to fool defenses with such consistency checks. Therefore, we present the first approach GLOW that copes with various attack requests by generating global layout-aware adversarial attacks, in which both categorical and geometric layout constraints are explicitly established. Specifically, we focus on object detection tasks and given a victim image, GLOW first localizes victim objects according to target labels. And then it generates multiple attack plans, together with their context-consistency scores. GLOW, on the one hand, is capable of handling various types of requests, including single or multiple victim objects, with or without specified victim objects. On the other hand, it produces a consistency score for each attack plan, reflecting the overall contextual consistency that both semantic category and global scene layout are considered. We conduct our experiments on MS COCO and Pascal. Extensive experimental results demonstrate that we can achieve about 30$\%$ average relative improvement compared to state-of-the-art methods in conventional single object attack request; Moreover, such superiority is also valid across more generic attack requests, under both white-box and zero-query black-box settings. Finally, we conduct comprehensive human analysis, which not only validates our claim further but also provides strong evidence that our evaluation metrics reflect human reviews well.

Learning fair representation in deep learning is essential to mitigate discriminatory outcomes and enhance trustworthiness. However, previous research has been commonly established on inappropriate assumptions prone to unrealistic counterfactuals and performance degradation. Although some proposed alternative approaches, such as employing correlation-aware causal graphs or proxies for mutual information, these methods are less practical and not applicable in general. In this work, we propose FAir DisEntanglement with Sensitive relevance (FADES), a novel approach that leverages conditional mutual information from the information theory perspective to address these challenges. We employ sensitive relevant code to direct correlated information between target labels and sensitive attributes by imposing conditional independence, allowing better separation of the features of interest in the latent space. Utilizing an intuitive disentangling approach, FADES consistently achieves superior performance and fairness both quantitatively and qualitatively with its straightforward structure. Specifically, the proposed method outperforms existing works in downstream classification and counterfactual generations on various benchmarks.

Federated learning (FL) has emerged as a new paradigm for privacy-preserving collaborative training. Under domain skew, the current FL approaches are biased and face two fairness problems. 1) Parameter Update Conflict: data disparity among clients leads to varying parameter importance and inconsistent update directions. These two disparities cause important parameters to potentially be overwhelmed by unimportant ones of dominant updates. It consequently results in significant performance decreases for lower-performing clients. 2) Model Aggregation Bias: existing FL approaches introduce unfair weight allocation and neglect domain diversity. It leads to biased model convergence objective and distinct performance among domains. We discover a pronounced directional update consistency in Federated Learning and propose a novel framework to tackle above issues. First, leveraging the discovered characteristic, we selectively discard unimportant parameter updates to prevent updates from clients with lower performance overwhelmed by unimportant parameters, resulting in fairer generalization performance. Second, we propose a fair aggregation objective to prevent global model bias towards some domains, ensuring that the global model continuously aligns with an unbiased model. The proposed method is generic and can be combined with other existing FL methods to enhance fairness. Comprehensive experiments on Digits and Office-Caltech demonstrate the high fairness and performance of our method.

The advances in the Neural Radiance Fields (NeRF) research offer extensive applications in diverse domains, but protecting their copyrights has not yet been researched in depth. Recently, NeRF watermarking has been considered one of the pivotal solutions for safely deploying NeRF-based 3D representations. However, existing methods are designed to apply only to implicit or explicit NeRF representations. In this work, we introduce an innovative watermarking method that can be employed in both representations of NeRF. This is achieved by fine-tuning NeRF to embed binary messages in the rendering process. In detail, we propose utilizing the discrete wavelet transform in the NeRF space for watermarking. Furthermore, we adopt a deferred back-propagation technique and introduce a combination with the patch-wise loss to improve rendering quality and bit accuracy with minimum trade-offs. We evaluate our method in three different aspects: capacity, invisibility, and robustness of the embedded watermarks in the 2D-rendered images. Our method achieves state-of-the-art performance with faster training speed over the compared state-of-the-art methods. Project page: https://kuai-lab.github.io/cvpr2024waterf/

Federated learning (FL) is a powerful technology that enables collaborative training of machine learning models without sharing private data among clients. The fundamental challenge in FL lies in learning over extremely heterogeneous data distributions, device capacities, and device state availabilities, all of which adversely impact performance and communication efficiency. While data heterogeneity has been well-studied in the literature, this paper introduces FLHetBench, the first FL benchmark targeted toward understanding device and state heterogeneity. FLHetBench comprises two new sampling methods to generate real-world device and state databases with varying heterogeneity and new metrics for quantifying the success of FL methods under these real-world constraints. Using FLHetBench, we conduct a comprehensive evaluation of existing methods and find that they struggle under these settings, which inspires us to propose BiasPrompt+, a new method employing staleness-aware aggregation and fast weights to tackle these new heterogeneity challenges. Experiments on various FL tasks and datasets validate the effectiveness of our BiasPrompt+ method and highlight the value of FLHetBench in fostering the development of more efficient and robust FL solutions under real-world device and state constraints.

Heterogeneous Federated Learning (HtFL) enables collaborative learning on multiple clients with different model architectures while preserving privacy. Despite recent research progress, knowledge sharing in HtFL is still difficult due to data and model heterogeneity. To tackle this issue, we leverage the knowledge stored in pre-trained generators and propose a new upload-efficient knowledge transfer scheme called Federated Knowledge-Transfer Loop (FedKTL). Our FedKTL can produce client-task-related prototypical image-vector pairs via the generator’s inference on the server. With these pairs, each client can transfer pre-existing knowledge from the generator to its local model through an additional supervised local task. We conduct extensive experiments on four datasets under two types of data heterogeneity with 14 kinds of models including CNNs and ViTs. Results show that our upload-efficient FedKTL surpasses seven state-of-the-art methods by up to 7.31% in accuracy. Moreover, our knowledge transfer scheme is applicable in scenarios with only one edge client. Code: https://github.com/TsingZ0/FedKTL

The modern surge in camera usage alongside widespread computer vision technology applications poses significant privacy and security concerns. Current artificial intelligence (AI) technologies aid in recognizing relevant events and assisting in daily tasks in homes, offices, hospitals, etc. The need to access or process personal information for these purposes raises privacy concerns. While software-level solutions like face de-identification provide a good privacy/utility trade-off, they present vulnerabilities to sniffing attacks. In this paper, we propose a hardware-level face de-identification method to solve this vulnerability. Specifically, our approach first learns an optical encoder along with a regression model to obtain a face heatmap while hiding the face identity from the source image. We also propose an anonymization framework that generates a new face using the privacy-preserving image, face heatmap, and a reference face image from a public dataset as input. We validate our approach with extensive simulations and hardware experiments.

Split Learning (SL) is a distributed learning framework renowned for its privacy-preserving features and minimal computational requirements. Previous research consistently highlights the potential privacy breaches in SL systems by server adversaries reconstructing training data. However, these studies often rely on strong assumptions or compromise system utility to enhance attack performance. This paper introduces a new semi-honest Data Reconstruction Attack on SL, named Feature-Oriented Reconstruction Attack (FORA). In contrast to prior works, FORA relies on limited prior knowledge, specifically that the server utilizes auxiliary samples from the public without knowing any client's private information. This allows FORA to conduct the attack stealthily and achieve robust performance. The key vulnerability exploited by FORA is the revelation of the model representation preference in the smashed data output by victim client. FORA constructs a substitute client through feature-level transfer learning, aiming to closely mimic the victim client's representation preference. Leveraging this substitute client, the server trains the attack model to effectively reconstruct private data. Extensive experiments showcase FORA's superior performance compared to state-of-the-art methods. Furthermore, the paper systematically evaluates the proposed method's applicability across diverse settings and advanced defense strategies.

Deep neural networks are known to be overconfident for what they don’t know in the wild, which is undesirable for decision-making in high-stakes applications. Despite quantities of existing works, most of them focus on detecting out-of-distribution (OOD) samples from unseen classes, while ignoring large parts of relevant failure sources like misclassified samples from known classes. In particular, recent studies reveal that prevalent OOD detection methods are actually harmful for misclassification detection (MisD), indicating that there seems to be a tradeoff between those two tasks. In this paper, we study the critical yet under-explored problem of unified failure detection, which aims to detect both misclassified and OOD examples. Concretely, we identify the failure of simply integrating learning objectives of misclassification and OOD detection, and show the potential of sequence learning. Inspired by this, we propose a reliable continual learning paradigm, whose spirit is to equip the model with MisD ability first, and then improve the OOD detection ability without degrading the already adequate MisD performance. Extensive experiments demonstrate that our method achieves strong unified failure detection performance. The code is available at https://github.com/Impression2805/RCL.

Prompt learning in pretrained visual-language models has shown remarkable flexibility across various downstream tasks. Leveraging its inherent lightweight nature, recent research attempted to integrate the powerful pretrained models into federated learning frameworks to simultaneously reduce communication costs and promote local training on insufficient data. Despite these efforts, current federated prompt learning methods lack specialized designs to systematically address severe data heterogeneities, e.g., data distribution with both label and feature shifts involved. To address this challenge, we present Federated Prompts Cooperation via Optimal Transport (FedOTP), which introduces efficient collaborative prompt learning strategies to capture diverse category traits on a per-client basis. Specifically, for each client, we learn a global prompt to extract consensus knowledge among clients, and a local prompt to capture client-specific category characteristics. Unbalanced Optimal Transport is then employed to align local visual features with these prompts, striking a balance between global consensus and local personalization. By relaxing one of the equality constraints, FedOTP enables prompts to focus solely on the core regions of image patches. Extensive experiments on datasets with various types of heterogeneities have demonstrated that our FedOTP outperforms the state-of-the-art methods.

Ethical concerns surrounding copyright protection and inappropriate content generation pose challenges for the practical implementation of diffusion models. One effective solution involves watermarking the generated images. However, existing methods often compromise the model performance or require additional training, which is undesirable for operators and users. To address this issue, we propose Gaussian Shading, a diffusion model watermarking technique that is both performance-lossless and training-free, while serving the dual purpose of copyright protection and tracing of offending content. Our watermark embedding is free of model parameter modifications and thus is plug-and-play. We map the watermark to latent representations following a standard Gaussian distribution, which is indistinguishable from latent representations obtained from the non-watermarked diffusion model. Therefore we can achieve watermark embedding with lossless performance, for which we also provide theoretical proof. Furthermore, since the watermark is intricately linked with image semantics, it exhibits resilience to lossy processing and erasure attempts. The watermark can be extracted by Denoising Diffusion Implicit Models (DDIM) inversion and inverse sampling. We evaluate Gaussian Shading on multiple versions of Stable Diffusion, and the results demonstrate that Gaussian Shading not only is performance-lossless but also outperforms existing methods in terms of robustness.

Large multi-modal models (LMMs) hold the potential to usher in a new era of automated visual assistance for people who are blind or low vision (BLV). Yet, these models have not been systematically evaluated on data captured by BLV users. We address this by empirically assessing CLIP, a widely-used LMM likely to underpin many assistive technologies. Testing 25 CLIP variants in a zero-shot classification task, we find that their accuracy is 15 percentage points lower on average for images captured by BLV users than web-crawled images. This disparity stems from CLIP's sensitivities to 1) image content (e.g. not recognizing disability objects as well as other objects); 2) image quality (e.g. not being robust to lighting variation); and 3) text content (e.g. not recognizing objects described by tactile adjectives as well as visual ones). We delve deeper with a textual analysis of three common pre-training datasets: LAION-400M, LAION-2B and DataComp-1B, showing that disability content is rarely mentioned. We then provide three examples that illustrate how the performance disparities extend to three downstream models underpinned by CLIP: OWL-ViT, CLIPSeg and DALL-E2. We find that few-shot learning with as few as 5 images can mitigate CLIP's quality-of-service disparities for BLV users in some scenarios, which we discuss alongside a set of other possible mitigations.

Model Inversion (MI) attacks aim to reconstruct private training data by abusing access to machine learning models. Contemporary MI attacks have achieved impressive attack performance posing serious threats to privacy. Meanwhile, all existing MI defense methods rely on regularization that has direct conflict with the training objective, resulting in noticeable degradation in model utility. In this work, we take a different perspective, and propose a novel and simple method based on transfer learning (TL) to render MI-robust models. Particularly, by leveraging TL, we limit the number of layers encoding sensitive information from private training dataset, thereby degrading the performance of MI attack. We conduct an analysis using Fisher Information to justify our method. Our defense is remarkably simple to implement. Without bells and whistles, we show in extensive experiments that our method achieves state-of-the-art (SOTA) MI robustness. Our code, pre-trained models, demo and inverted data are included in Supp.

Deep Neural Networks (DNNs) are powerful tools for various computer vision tasks, yet they often struggle with reliable uncertainty quantification — a critical requirement for real-world applications. Bayesian Neural Networks (BNN) are equipped for uncertainty estimation but cannot scale to large DNNs where they are highly instable to train.To address this challenge, we introduce the Adaptable Bayesian Neural Network (ABNN), a simple and scalable strategy to seamlessly transform DNNs into BNNs in a post-hoc manner with minimal computational and training overheads.ABNN preserves main predictive properties of DNNs while enhancing their uncertainty quantification abilities through simple BNN adaptation layers (attached to normalization layers) and few fine-tuning steps on pre-trained models. We conduct extensive experiments across multiple datasets for image classification and semantic segmentation tasks, and our results demonstrate that ABNN achieves state-of-the-art performance without the computational budget typically associated with ensemble methods.

The widespread use of cloud-based face recognition technology raises privacy concerns, as unauthorized access to face images can expose personal information or be exploited for fraudulent purposes. In response, privacy-preserving face recognition (PPFR) schemes have emerged to hide visual information and thwart unauthorized access. However, the validation methods employed by these schemes often rely on unrealistic assumptions, leaving doubts about their true effectiveness in safeguarding facial privacy. In this paper, we introduce a new approach to privacy validation called Minimum Assumption Privacy Protection Validation (Map$^2$V). This is the first exploration of formulating a privacy validation method utilizing deep image priors and zeroth-order gradient estimation, with the potential to serve as a general framework for PPFR evaluation. Building upon Map$^2$V, we comprehensively validate the privacy-preserving capability of PPFRs through a combination of human and machine vision. The experiment results and analysis demonstrate the effectiveness and generalizability of the proposed Map$^2$V, showcasing its superiority over native privacy validation methods from PPFR works of literature. Additionally, this work exposes privacy vulnerabilities in evaluated state-of-the-art PPFR schemes, laying the foundation for the subsequent effective proposal of countermeasures. The source code is available at https://github.com/Beauty9882/MAP2V.

The unauthorized use of personal data for commercial purposes and the covert acquisition of private data for training machine learning models continue to raise concerns. To address these issues, researchers have proposed availability attacks that aim to render data unexploitable. However, many availability attack methods can be easily disrupted by adversarial training. Although some robust methods can resist adversarial training, their protective effects are limited. In this paper, we re-examine the existing availability attack methods and propose a novel two-stage min-max-min optimization paradigm to generate robust unlearnable noise. The inner min stage is utilized to generate unlearnable noise, while the outer min-max stage simulates the training process of the poisoned model. Additionally, we formulate the attack effects and use it to constrain the optimization objective. Comprehensive experiments have revealed that the noise generated by our method can lead to a decline in test accuracy for adversarially trained poisoned models by up to approximately $30\%$, in comparison to state-of-the-art methods.

Text-to-image generative models are becoming increasingly popular and accessible to the general public. As these models see large-scale deployments, it is necessary to deeply investigate their safety and fairness to not disseminate and perpetuate any kind of biases. However, existing works focus on detecting closed sets of biases defined a priori, limiting the studies to well-known concepts. In this paper, we tackle the challenge of open-set bias detection in text-to-image generative models presenting OpenBias, a new pipeline that identifies and quantifies the severity of biases agnostically, without access to any precompiled set. OpenBias has three stages. In the first phase, we leverage a Large Language Model (LLM) to propose biases given a set of captions. Secondly, the target generative model produces images using the same set of captions. Lastly, a Vision Question Answering model recognizes the presence and extent of the previously proposed biases. We study the behavior of Stable Diffusion 1.5, 2, and XL emphasizing new biases, never investigated before. Via quantitative experiments, we demonstrate that OpenBias agrees with current closed-set bias detection methods and human judgement.

To alleviate the utility degradation of deep learning image classification with differential privacy (DP), employing extra public data or pre-trained models has been widely explored. Recently, the use of in-distribution public data has been investigated, where a tiny subset of data owners share their data publicly. In this paper, we investigate a framework that leverages recent diffusion models to amplify the information of public data. Subsequently, we identify data diversity and generalization gap between public and private data as critical factors addressing the limited size of public data. While assuming 4\% of training data as public, our method achieves 85.48\% on CIFAR-10 without using pre-trained models, with a privacy budget of $(2,10^{-5})$.

Federated learning is a decentralized learning paradigm introduced to preserve privacy of client data. Despite this, prior work has shown that an attacker at the server can still reconstruct the private training data using only the client updates. These attacks are known as data reconstruction attacks and fall into two major categories: gradient inversion (GI) and linear layer leakage attacks (LLL). However, despite demonstrating the effectiveness of these attacks in breaching privacy, prior work has not investigated the usefulness of the reconstructed data for downstream tasks. In this work, we explore data reconstruction attacks through the lens of training and improving models with leaked data. We demonstrate the effectiveness of both GI and LLL attacks in maliciously training models using the leaked data more accurately than a benign federated learning strategy. Counter-intuitively, this bump in training quality can occur despite limited reconstruction quality or a small total number of leaked images. Finally, we show the limitations of these attacks for downstream training, individually for GI attacks and for LLL attacks.

State-of-the-art personalized text-to-image generation systems are usually trained on a few reference images to learn novel visual representations. However, this is likely to incur infringement of copyright for the reference image owners, when these images are personal and publicly available. Recent progress has been made in protecting these images from unauthorized use by adding protective noises. Yet current protection methods work under the assumption that these protected images are not changed, which is in contradiction to the fact that most public platforms intend to modify user-uploaded content, e.g., image compression. This paper introduces a robust watermarking method, namely InMark, to protect images from unauthorized learning. Inspired by influence functions, the proposed method forges protective watermarks on more important pixels for these reference images from both heuristic and statistical perspectives. In this way, the personal semantics of these images are under protection even if these images are modified to some extent. Extensive experiments demonstrate that the proposed InMark outperforms previous state-of-the-art methods in both protective performance and robustness.

Despite the remarkable success of Vision Transformers (ViT) across diverse fields in computer vision, they have a clear drawback of expensive adaption cost for downstream tasks due to the increased scale. To address this, Visual Prompt Tuning (VPT) incorporates learnable parameters in the input space of ViT. While freezing the ViT backbone and tuning only the prompts, it exhibits superior performances to full fine-tuning. However, despite the outstanding advantage, we point out that VPT may lead to serious unfairness in downstream classification. Initially, we investigated the causes of unfairness in VPT, identifying the biasedly pre-trained ViT as a principal factor. Motivated by this observation, we propose a Fair Visual Prompt Tuning (Fair-VPT) which removes biased information in the pre-trained ViT while adapting it to downstream classification tasks. To this end, we categorize prompts into "cleaner prompts" and "target prompts''. Based on this, we encode the class token in two different ways by either masking or not masking the target prompts in the self-attention process. These encoded tokens are trained with distinct objective functions, resulting in the inclusion of different information in the target and cleaner prompts. Moreover, we introduce a disentanglement loss based on contrastive learning to further decorrelate them. In experiments across diverse benchmarks, the proposed method demonstrates the most superior performance in terms of balanced classification accuracy and fairness.

We propose a novel contrastive learning framework to effectively address the challenges of data heterogeneity in federated learning. We first analyze the inconsistency of gradient updates across clients during local training and establish its dependence on the distribution of feature representations, leading to the derivation of the supervised contrastive learning (SCL) objective to mitigate local deviations.In addition, we show that a na\"ive adoption of SCL leads to representation collapse, resulting in slow convergence and limited performance gains. To address this issue, we introduce a relaxed contrastive learning loss that imposes a divergence penalty on excessively similar sample pairs within each class. This strategy prevents collapsed representations and enhances feature transferability, facilitating collaborative training and leading to significant performance improvements. Our framework outperforms all existing federated learning approaches by huge margins on the standard benchmarks through extensive experimental results. We plan to release the source code of our work for better reproducibility.

Fairness is a critical concern in deep learning, especially in healthcare, where these models influence diagnoses and treatment decisions. Although fairness has been investigated in the vision-only domain, the fairness of medical vision-language (VL) models remains unexplored due to the scarcity of medical VL datasets for studying fairness. To bridge this research gap, we introduce the first fair vision-language medical dataset FairVLMed that provides detailed demographic attributes, ground-truth labels, and clinical notes to facilitate an in-depth examination of fairness within VL foundation models. Using FairVLMed, we conduct a comprehensive fairness analysis of two widely-used VL models (CLIP and BLIP2), pre-trained on both natural and medical domains, across four different protected attributes. Our results highlight significant biases in all VL models, with Asian, Male, Non-Hispanic, and Spanish being the preferred subgroups across the protected attributes of race, gender, ethnicity, and language, respectively. In order to alleviate these biases, we propose FairCLIP, an optimal-transport-based approach that achieves a favorable trade-off between performance and fairness by reducing the Sinkhorn distance between the overall sample distribution and the distributions corresponding to each demographic group. As the first VL dataset of its kind, FairVLMed holds the potential to catalyze advancements in the development of machine learning models that are both ethically aware and clinically effective. Our dataset and code are available at https://ophai.hms.harvard.edu/datasets/fairvlmed10k.

Ensuring the legal usage of deep models is crucial to promoting trustable, accountable, and responsible artificial intelligence innovation. Current passport-based methods that obfuscate model functionality for license-to-use and ownership verifications suffer from capacity and quality constraints, as they require retraining the owner model for new users. They are also vulnerable to advanced Expanded Residual Block ambiguity attacks. We propose Steganographic Passport, which uses an invertible steganographic network to decouple license-to-use from ownership verification by hiding the user’s identity images into the owner-side passport and recovering them from their respective user-side passports. An irreversible and collision-resistant hash function is used to avoid exposing the owner-side passport from the derived user-side passports and increase the uniqueness of the model signature. To safeguard both the passport and model’s weights against advanced ambiguity attacks, an activation-level obfuscation is proposed for the verification branch of the owner’s model. By jointly training the verification and deployment branches, their weights become tightly coupled. The proposed method supports agile licensing of deep models by providing a strong ownership proof and license accountability without requiring a separate model retraining for the admission of every new user. Experiment results show that our Steganographic Passport outperforms other passport-based deep model protection methods in robustness against various known attacks.

In Federated Learning (FL), the issue of statistical data heterogeneity has been a significant challenge to the field's ongoing development. This problem is further exacerbated when clients' data vary in modalities. In response to these issues of statistical heterogeneity and modality incompatibility, we propose the Adaptive Hyper-graph Aggregation framework, a novel solution for Modality-Agnostic Federated Learning. We design a Modular Architecture for Local Model with single modality, setting the stage for efficient intra-modality sharing and inter-modality complementarity. An innovative Global Consensus Prototype Enhancer is crafted to assimilate and broadcast global consensus knowledge within the network. At the core of our approach lies the Adaptive Hyper-graph Learning Strategy, which effectively tackles the inherent challenges of modality incompatibility and statistical heterogeneity within federated learning environments, accomplishing this adaptively even without the server being aware of the clients' modalities. Our approach, tested on three multimodal benchmark datasets, demonstrated strong performance across diverse data distributions, affirming its effectiveness in multimodal federated learning. The code will be made available publicly at https://anonymous.4open.science/r/HAMFL.

Recent studies have noted an intriguing phenomenon termed Neural Collapse, that is, when the neural networks establish the right correlation between feature spaces and the training targets, their last-layer features, together with the classifier weights, will collapse into a stable and symmetric structure. In this paper, we extend the investigation of Neural Collapse to the biased datasets with imbalanced attributes. We observe that models will easily fall into the pitfall of shortcut learning and form a biased, non-collapsed feature space at the early period of training, which is hard to reverse and limits the generalization capability. To tackle the root cause of biased classification, we follow the recent inspiration of prime training, and propose an avoid-shortcut learning framework without additional training complexity. With well-designed shortcut primes based on Neural Collapse structure, the models are encouraged to skip the pursuit of simple shortcuts and naturally capture the intrinsic correlations. Experimental results demonstrate that our method induces a better convergence property during training, and achieves state-of-the-art generalization performance on both synthetic and real-world biased datasets.

In the image classification task, deep neural networks frequently rely on bias attributes that are spuriously correlated with a target class in the presence of dataset bias, resulting in degraded performance when applied to data without bias attributes.The task of debiasing aims to compel classifiers to learn intrinsic attributes that inherently define a target class rather than focusing on bias attributes.While recent approaches mainly focus on emphasizing the learning of data samples without bias attributes (i.e., bias-conflicting samples) compared to samples with bias attributes (i.e., bias-aligned samples), they fall short of directly guiding models where to focus for learning intrinsic features.To address this limitation, this paper proposes a method that provides the model with explicit spatial guidance that indicates the region of intrinsic features. We first identify the intrinsic features by investigating the class-discerning common features between a bias-aligned (BA) sample and a bias-conflicting (BC) sample (i.e., bias-contrastive pair).Next, we enhance the intrinsic features in the BA sample that are relatively under-exploited for prediction compared to the BC sample. To construct the bias-contrastive pair without using bias information, we introduce a bias-negative score that distinguishes BC samples from BA samples employing a biased model.The experiments demonstrate that our method achieves state-of-the-art performance on synthetic and real-world datasets with various levels of bias severity.

Neural network pruning, particularly channel pruning, is a widely used technique for compressing deep learning models to enable their deployment on edge devices with limited resources. Typically, redundant weights or structures are removed to achieve the target resource budget. Although data-driven pruning approaches have proven to be more effective, they cannot be directly applied to federated learning (FL), which has emerged as a popular technique in edge computing applications, because of distributed and confidential datasets. In response to this challenge, we design a new network pruning method for FL. We propose device-wise sub-networks for each device, assuming that the data distribution is similar within each device. These sub-networks are generated through sub-network embeddings and a hypernetwork. To further minimize memory usage and communication costs, we permanently prune the full model to remove weights that are not useful for all devices. During the FL process, we simultaneously train the device-wise sub-networks and the base sub-network to facilitate the pruning process. We then finetune the pruned model with device-wise sub-networks to regain performance. In addition, we provided the theoretical guarantee of convergence for our method. Our method achieves better performance and resource trade-off than other well-established network pruning baselines, as demonstrated through extensive experiments on CIFAR-10, CIFAR-100, and TinyImageNet.

Data privacy is of great concern in cloud machine-learning service platforms, when sensitive data are exposed to service providers. While private computing environments (e.g., secure enclaves), and cryptographic approaches (e.g., homomorphic encryption) provide strong privacy protection, their computing performance still falls short compared to cloud GPUs. To achieve privacy protection with high computing performance, we propose Delta, a new private training and inference framework, with comparable model performance as non-private centralized training. Delta features two asymmetric data flows: the main information-sensitive flow and the residual flow. The main part flows into a small model while the residuals are offloaded to a large model. Specifically, Delta embeds the information-sensitive representations into a low-dimensional space while pushing the information-insensitive part into high-dimension residuals. To ensure privacy protection, the low-dimensional information-sensitive part is secured and fed to a small model in a private environment. On the other hand, the residual part is sent to fast cloud GPUs, and processed by a large model. To further enhance privacy and reduce the communication cost, Delta applies a random binary quantization technique along with a DP-based technique to the residuals before sharing them with the public platform. We theoretically show that Delta guarantees differential privacy in the public environment and greatly reduces the complexity in the private environment. We conduct empirical analyses on CIFAR-10, CIFAR-100 and ImageNet datasets and ResNet-18 and ResNet-34, showing that Delta achieves strong privacy protection, fast training, and inference without significantly compromising the model utility.

The booming use of text-to-image generative models has raised concerns about their high risk of producing copyright-infringing content. While probabilistic copyright protection methods provide a probabilistic guarantee against such infringement, in this paper, we introduce Virtually Assured Amplification Attack (VA3), a novel online attack framework that exposes the vulnerabilities of these protection mechanisms. The proposed framework significantly amplifies the probability of generating infringing content on the sustained interactions with generative models and a non-trivial lower-bound on the success probability of each engagement. Our theoretical and experimental results demonstrate the effectiveness of our approach under various scenarios. These findings highlight the potential risk of implementing probabilistic copyright protection in practical applications of text-to-image generative models.

Retrieval Augmented Generation (RAG) is emerging as a flexible and robust technique to adapt models to private users data without training, to handle credit attribution, and to allow efficient machine unlearning at scale. However, RAG techniques for image generation may lead to parts of the retrieved samples being copied in the model's output. To reduce risks of leaking private information contained in the retrieved set, we introduce Copy-Protected generation with Retrieval (CPR), a new method for RAG with strong copyright protection guarantees in a mixed-private setting for diffusion models. CPR allows to condition the output of diffusion models on a set of retrieved images, while also guaranteeing that unique identifiable information about those example is not exposed in the generated outputs. In particular, it does so by sampling from a mixture of public (safe) distribution and private (user) distribution by merging their diffusion scores at inference.We prove that CPR satisfies Near Access Freeness (NAF) which bounds the amount of information an attacker may be able to extract from the generated images. We provide two algorithms for copyright protection, CPR-KL and CPR-Choose. Unlike previously proposed rejection-sampling-based NAF methods, our methods enable efficient copyright-protected sampling with a single run of backward diffusion. We show that our method can be applied to any pre-trained conditional diffusion model, such as Stable Diffusion or unCLIP. In particular, we empirically show that applying CPR on top of unCLIP improves quality and text-to-image alignment of the generated results (81.4 to 83.17 on TIFA benchmark), while enabling credit attribution, copy-right protection, and deterministic, constant time, unlearning.

Federated learning often suffers from slow and unstable convergence due to the heterogeneous characteristics of participating client datasets.Such a tendency is aggravated when the client participation ratio is low since the information collected from the clients has large variations.To address this challenge, we propose a simple but effective federated learning framework, which improves the consistency across clients and facilitates the convergence of the server model.This is achieved by making the server broadcast a global model with a gradient acceleration.This strategy enables the proposed approach to convey the projective global update information to participants effectively without additional client memory for storing previous models, and extra communication costs.We also regularize local updates by aligning each client with the overshot global model to reduce bias and improve the stability of our algorithm.We provide the theoretical convergence rate of our algorithm and demonstrate remarkable performance gains in terms of accuracy and communication efficiency compared to the state-of-the-art methods, especially with low client participation rates.We plan to release our code to facilitate the reproduction of our work.

Spurious correlations can cause strong biases in deep neural networks, impairing generalization ability. While most existing debiasing methods require full supervision on either spurious attributes or target labels, training a debiased model from a limited amount of both annotations is still an open question. To address this issue, we investigate an interesting phenomenon using the spectral analysis of latent representations: spuriously correlated attributes make neural networks inductively biased towards encoding lower effective rank representations. We also show that a rank regularization can amplify this bias in a way that encourages highly correlated features. Leveraging these findings, we propose a self-supervised debiasing framework potentially compatible with unlabeled samples. Specifically, we first pretrain a biased encoder in a self-supervised manner with the rank regularization, serving as a semantic bottleneck to enforce the encoder to learn the spuriously correlated attributes. This biased encoder is then used to discover and upweight bias-conflicting samples in a downstream task, serving as a boosting to effectively debias the main model. Remarkably, the proposed debiasing framework significantly improves the generalization performance of self-supervised learning baselines and, in some cases, even outperforms state-of-the-art supervised debiasing approaches.

The unprecedented capture and application of face images raise increasing concerns on anonymization to fight against privacy disclosure. Most existing methods may suffer from the problem of excessive change of the identity-independent information or insufficient identity protection. In this paper, we present a new face anonymization approach by distracting the intrinsic and extrinsic identity attentions. On the one hand, we anonymize the identity information in the feature space by distracting the intrinsic identity attention. On the other, we anonymize the visual clues (i.e. appearance and geometry structure) by distracting the extrinsic identity attention. Our approach allows for flexible and intuitive manipulation of face appearance and geometry structure to produce diverse results, and it can also be used to instruct users to perform personalized anonymization. We conduct extensive experiments on multiple datasets and demonstrate that our approach outperforms state-of-the-art methods.

Unsupervised (US) video anomaly detection (VAD) in surveillance applications is gaining more popularity lately due to its practical real-world applications. Due to the extremely challenging nature of this task, where learning is carried out without any annotations, privacy-critical collaborative learning of US-VAD systems has not been studied yet. As surveillance videos are privacy sensitive and the availability of large-scale video data may enable better US-VAD systems, collaborative learning can be highly rewarding in this setting. In this paper, we propose a new baseline for anomaly detection capable of localizing anomalous events in complex surveillance scenarios in a fully unsupervised fashion without any labels on a privacy-retaining participant-based distributed training configuration. Additionally, we propose three new evaluation protocols to extensively evaluate anomaly detection approaches on various scenarios of collaborations and data availability. Moreover, based on these protocols, we modify existing VAD datasets to extensively evaluate our approach as well as existing US SOTA methods on two large-scale datasets including UCF-Crime and XD-Violence. All proposed evaluation protocols, dataset splits, and codes are available here: \href{https://github.com/AnasEmad11/CLAP}{https://github.com/AnasEmad11/CLAP}.

Deep neural networks are prone to capture correlations between spurious attributes and class labels, leading to low accuracy on some combinations of class labels and spurious attribute values. When a spurious attribute represents a protected class, these low-accuracy groups can manifest discriminatory bias. Existing methods attempting to improve worst-group accuracy assume the training data, validation data, or both are reliably labeled by the spurious attribute. But a model may be perceived to be biased towards a concept that is not represented by pre-existing labels on the training data. In these situations, the spurious attribute must be defined with external information. We propose Concept Correction, a framework that represents a concept as a curated set of images from any source, then labels each training sample by its similarity to the concept set to control spurious correlations. For example, concept sets representing gender can be used to measure and control gender bias even without explicit labels. We demonstrate and evaluate an instance of the framework as Concept DRO, which uses concept sets to estimate group labels, then uses these labels to train with a state of the art distributively robust optimization objective. We show that Concept DRO outperforms existing methods that do not require labels of spurious attributes by up to 33.1% on three image classification datasets and is competitive with the best methods that assume access to labels. We consider how the size and quality of the concept set influences performance and find that even smaller, manually curated sets of noisy AI-generated images are effective at controlling spurious correlations, suggesting that high-quality, reusable concept sets are easy to create and effective in reducing bias.

Anomaly detection (AD) aims to identify defective images and localize their defects (if any). Ideally, AD models should be able to detect defects over many image classes; without relying on hard-coded class names that can be uninformative or inconsistent across datasets; learn without anomaly supervision; and be robust to the long-tailed distributions of real-world applications. To address these challenges, we formulate the problem of long-tailed AD by introducing several datasets with different levels of class imbalance and metrics for performance evaluation. We then propose a novel method, LTAD, to detect defects from multiple and long-tailed classes, without relying on dataset class names. LTAD combines AD by reconstruction and semantic AD modules. AD by reconstruction is implemented with a transformer-based reconstruction module. Semantic AD is implemented with a binary classifier, which relies on learned pseudo class names and a pretrained foundation model. These modules are learned over two phases. Phase 1 learns the pseudo-class names and a variational autoencoder (VAE) for feature synthesis that augments the training data to combat long-tails. Phase 2 then learns the parameters of the reconstruction and classification modules of LTAD. Extensive experiments using the proposed long-tailed datasets show that LTAD substantially outperforms the state-of-the-art methods for most forms of dataset imbalance. The long-tailed dataset split is available at https://zenodo.org/records/10854201

Context-aware emotion recognition (CAER) has recently boosted the practical applications of affective computing techniques in unconstrained environments. Mainstream CAER methods invariably extract ensemble representations from diverse contexts and subject-centred characteristics to perceive the target person's emotional state. Despite advancements, the biggest challenge remains due to context bias interference. The harmful bias forces the models to rely on spurious correlations between background contexts and emotion labels in likelihood estimation, causing severe performance bottlenecks and confounding valuable context priors. In this paper, we propose a counterfactual emotion inference (CLEF) framework to address the above issue. Specifically, we first formulate a generalized causal graph to decouple the causal relationships among the variables in CAER. Following the causal graph, CLEF introduces a non-invasive context branch to capture the adverse direct effect caused by the context bias. During the inference, we eliminate the direct context effect from the total causal effect by comparing factual and counterfactual outcomes, resulting in bias mitigation and robust prediction.As a model-agnostic framework, CLEF can be readily integrated into existing methods, bringing consistent performance gains.

Multimodal sentiment analysis (MSA) aims to understand human sentiment through multimodal data. Most MSA efforts are based on the assumption of modality completeness. However, in real-world applications, some practical factors cause uncertain modality missingness, which drastically degrades the model's performance. To this end, we propose a Correlation-decoupled Knowledge Distillation (CorrKD) framework for the MSA task under uncertain missing modalities. Specifically, we present a sample-level contrastive distillation mechanism that transfers comprehensive knowledge containing cross-sample correlations to reconstruct missing semantics. Moreover, a category-guided prototype distillation mechanism is introduced to capture cross-category correlations using category prototypes to align feature distributions and generate favorable joint representations. Eventually, we design a response-disentangled consistency distillation strategy to optimize the sentiment decision boundaries of the student network through response disentanglement and mutual information maximization. Comprehensive experiments on three datasets indicate that our framework can achieve favorable improvements compared with several baselines.

Denoising diffusion probabilistic models (DDPMs) employ a sequence of white Gaussian noise samples to generate an image. In analogy with GANs, those noise maps could be considered as the latent code associated with the generated image. However, this native noise space does not possess a convenient structure, and is thus challenging to work with in editing tasks. Here, we propose an alternative latent noise space for DDPM that enables a wide range of editing operations via simple means, and present an inversion method for extracting these edit-friendly noise maps for any given image (real or synthetically generated). As opposed to the native DDPM noise space, the edit-friendly noise maps do not have a standard normal distribution and are not statistically independent across timesteps. However, they allow perfect reconstruction of any desired image, and simple transformations on them translate into meaningful manipulations of the output image (e.g. shifting, color edits). Moreover, in text-conditional models, fixing those noise maps while changing the text prompt, modifies semantics while retaining structure. We illustrate how this property enables text-based editing of real images via the diverse DDPM sampling scheme (in contrast to the popular non-diverse DDIM inversion). We also show how it can be used within existing diffusion-based editing methods to improve their quality and diversity.

Advances in camera-based physiological monitoring have enabled the robust, non-contact measurement of respiration and the cardiac pulse, which are known to be indicative of the sleep stage. This has led to research into camera-based sleep monitoring as a promising alternative to ``gold-standard'' polysomnography, which is cumbersome, expensive to administer, and hence unsuitable for longer-term clinical studies. In this paper, we introduce SleepVST, a transformer model which enables state-of-the-art performance in camera-based sleep stage classification (sleep staging). After pre-training on contact sensor data, SleepVST outperforms existing methods for cardio-respiratory sleep staging on the SHHS and MESA datasets, achieving total Cohen's kappa scores of 0.75 and 0.77 respectively. We then show that SleepVST can be successfully transferred to cardio-respiratory waveforms extracted from video, enabling fully contact-free sleep staging. Using a video dataset of 50 nights, we achieve a total accuracy of 78.8\% and a Cohen's $\kappa$ of 0.71 in four-class video-based sleep staging, setting a new state-of-the-art in the domain.

A handful of visual foundation models (VFMs) have recently emerged as the backbones for numerous downstream tasks. VFMs like CLIP, DINOv2, SAM are trained with distinct objectives, exhibiting unique characteristics for various downstream tasks. We find that despite their conceptual differences, these models can be effectively merged into a unified model through multi-teacher distillation. We name this approach AM-RADIO (Agglomerative Model -- Reduce All Domains Into One). This integrative approach not only surpasses the performance of individual teacher models but also amalgamates their distinctive features, such as zero-shot vision-language comprehension, detailed pixel-level understanding, and open vocabulary segmentation capabilities. In pursuit of the most hardware-efficient backbone, we evaluated numerous architectures in our multi-teacher distillation pipeline using the same training recipe. This led to the development of a novel architecture (E-RADIO) that exceeds the performance of its predecessors and is at least 7x faster than the teacher models. Our comprehensive benchmarking process covers downstream tasks including ImageNet classification, ADE20k semantic segmentation, COCO object detection and LLaVa-1.5 framework.

We introduce a new task -- language-driven video inpainting, which uses natural language instructions to guide the inpainting process. This approach overcomes the limitations of traditional video inpainting methods that depend on manually labeled binary masks, a process often tedious and labor-intensive. We present the Remove Objects from Videos by Instructions (ROVI) dataset, containing 5,650 videos and 9,091 inpainting results, to support training and evaluation for this task. We also propose a novel diffusion-based language-driven video inpainting framework, the first end-to-end baseline for this task, integrating Multimodal Large Language Models to understand and execute complex language-based inpainting requests effectively. Our comprehensive results showcase the dataset's versatility and the model's effectiveness in various language-instructed inpainting scenarios. We have made datasets, code, and models publicly available at \url{https://github.com/jianzongwu/Language-Driven-Video-Inpainting}.

Federated Learning (FL) aggregates locally trained models from individual clients to construct a global model. While FL enables learning a model with data privacy, it often suffers from significant performance degradation when clients have heterogeneous data distributions. This data heterogeneity causes the model to forget the global knowledge acquired from previously sampled clients after being trained on local datasets. Although the introduction of proximal objectives in local updates helps to preserve global knowledge, it can also hinder local learning by interfering with local objectives. To address this problem, we propose a novel method, Federated Stabilized Orthogonal Learning (FedSOL), which adopts an orthogonal learning strategy to balance the two conflicting objectives. FedSOL is designed to identify gradients of local objectives that are inherently orthogonal to directions affecting the proximal objective. Specifically, FedSOL targets parameter regions where learning on the local objective is minimally influenced by proximal weight perturbations. Our experiments demonstrate that FedSOL consistently achieves state-of-the-art performance across various scenarios.

We present UnionFormer, a novel framework that integrates tampering clues across three views by unified learning for image manipulation detection and localization. Specifically, we construct a BSFI-Net to extract tampering features from RGB and noise views, achieving enhanced responsiveness to boundary artifacts while modulating spatial consistency at different scales. Additionally, to explore the inconsistency between objects as a new view of clues, we combine object consistency modeling with tampering detection and localization into a three-task unified learning process, allowing them to promote and improve mutually. Therefore, we acquire a unified manipulation discriminative representation under multi-scale supervision that consolidates information from three views. This integration facilitates highly effective concurrent detection and localization of tampering. We perform extensive experiments on diverse datasets, and the results show that the proposed approach outperforms state-of-the-art methods in tampering detection and localization.

Motion blur is a frequently observed image artifact, especially under insufficient illumination where exposure time has to be prolonged so as to collect more photons for a bright enough image. Rather than simply removing such blurring effects, recent research has aimed at decomposing a blurry image into multiple sharp images with spatial and temporal coherence. Since motion blur decomposition itself is highly ambiguous, priors from neighbouring frames or human annotation are usually needed for motion disambiguation. In this paper, inspired by the complementary exposure characteristics of a global shutter (GS) camera and a rolling shutter (RS) camera, we propose to utilize the ordered scanline-wise delay in a rolling shutter image to robustify motion decomposition of a single blurry image. To evaluate this novel dual imaging setting, we construct a triaxial system to collect realistic data, as well as a deep network architecture that explicitly addresses temporal and contextual information through reciprocal branches for cross-shutter motion blur decomposition. Experiment results have verified the effectiveness of our proposed algorithm, as well as the validity of our dual imaging setting.

Once only a few-shot annotated samples are available, the performance of learning-based object detection would be heavily dropped. Many few-shot object detection (FSOD) methods have been proposed to tackle this issue by adopting image-level augmentations in linear manners. Nevertheless, those handcrafted enhancements often suffer from limited diversity and lack of semantic awareness, resulting in unsatisfactory performance. To this end, we propose a Semantic-guided Non-linear Instance-level Data Augmentation method (SNIDA) for FSOD by decoupling the foreground and background to increase their diversities respectively. We design a semantic awareness enhancement strategy to separate objects from backgrounds. Concretely, masks of instances are extracted by an unsupervised semantic segmentation module. Then the diversity of samples would be improved by fusing instances into different backgrounds. Considering the shortcomings of augmenting images in a limited transformation space of existing traditional data augmentation methods, we introduce an object reconstruction enhancement module. The aim of this module is to generate sufficient diversity and non-linear training data at the instance level through a semantic-guided masked autoencoder. In this way, the potential of data can be fully exploited in various object detection scenarios. Extensive experiments on PASCAL VOC and MS-COCO demonstrate that the proposed method outperforms baselines by a large margin and achieves new state-of-the-art results under different shot settings.

With the emergence of AR/VR, 3D models are in tremendous demand. However, conventional 3D modeling with Computer-Aided Design software requires much expertise and is difficult for novice users. We find that AR/VR devices, in addition to serving as effective display mediums, can offer a promising potential as an intuitive 3D model creation tool, especially with the assistance of AI generative models. Here, we propose Deep3DVRSketch, the first 3D model generation network that inputs 3D VR sketches from novice users and generates highly consistent 3D models in multiple categories within seconds, irrespective of the users’ drawing abilities. We also contribute KO3D+, the largest 3D sketch-shape dataset. Our method pre-trains a conditional diffusion model on quality 3D data, then finetunes an encoder to map 3D sketches onto the generator’s manifold using an adaptive curriculum strategy for limited ground truths. In our experiment, our approach achieves state-of-the-art performance in both model quality and fidelity with real-world input from novice users, and users can even draw and obtain very detailed geometric structures. In our user study, users were able to complete the 3D modeling tasks over 10 times faster using our approach compared to conventional CAD software tools. We believe that our Deep3DVRSketch and KO3D+ dataset can offer a promising solution for future 3D modeling in metaverse era. Check the project page at http://research.kokoni3d.com/Deep3DVRSketch.

We propose SketchINR, to advance the representation of vector sketches with implicit neural models. A variable length vector sketch is compressed into a latent space of fixed dimension that implicitly encodes the underlying shape as a function of time and strokes. The learned function predicts the $xy$ point coordinates in a sketch at each time and stroke. Despite its simplicity, SketchINR outperforms existing representations at multiple tasks: (i) Encoding an entire sketch dataset into a fixed size latent vector, SketchINR gives $60\times$ and $10\times$ data compression over raster and vector sketches, respectively. (ii) SketchINR's auto-decoder provides a much higher-fidelity representation than other learned vector sketch representations, and is uniquely able to scale to complex vector sketches such as FS-COCO. (iii) SketchINR supports parallelisation that can decode/render $\sim$$100\times$ faster than other learned vector representations such as SketchRNN. (iv) SketchINR, for the first time, emulates the human ability to reproduce a sketch with varying abstraction in terms of number and complexity of strokes. As a first look at implicit sketches, SketchINR's compact high-fidelity representation will support future work in modelling long and complex sketches.

Collaborative perception enhances perception performance by enabling autonomous vehicles to exchange complementary information. Despite its potential to revolutionize the mobile industry, challenges in various environments, such as communication bandwidth limitations, localization errors and information aggregation inefficiencies, hinder its implementation in practical applications.In this work, we propose ERMVP, a communication-Efficient and collaboration-Robust Multi-Vehicle Perception method in challenging environments. Specifically, ERMVP has three distinct strengths:i) It utilizes the hierarchical feature sampling strategy to abstract a representative set of feature vectors, using less communication overhead for efficient communication;ii) It employs the sparse consensus features to execute precise spatial location calibrations, effectively mitigating the implications of vehicle localization errors;iii) A pioneering feature fusion and interaction paradigm is introduced to integrate holistic spatial semantics among different vehicles and data sources.To thoroughly validate our method, we conduct extensive experiments on real-world and simulated datasets. The results demonstrate that the proposed ERMVP is significantly superior to the state-of-the-art collaborative perception methods.

Multimodal learning has advanced the performance for many vision-language tasks. However, most existing works in embodied dialog research focus on navigation and leave the localization task understudied. The few existing dialog-based localization approaches assume the availability of entire dialog prior to localizaiton, which is impractical for deployed dialog-based localization. In this paper, we propose DiaLoc, a new dialog-based localization framework which aligns with a real human operator behavior. Specifically, we produce an iterative refinement of location predictions which can visualize current pose believes after each dialog turn. DiaLoc effectively utilizes the multimodal datafor multi-shot localization, where a fusion encoder fuses vision and dialog information iteratively. We achieve state-of-the-art results on embodied dialog-based localization task, in single-shot (+7.08% in Acc5@valUnseen) and multi-shot settings (+10.85% in Acc5@valUnseen). DiaLoc narrows the gap between simulation and real-world applications, opening doors for future research on collaborative localization and navigation.

We introduce WildlifeMapper (WM), a flexible model designed to detect, locate, and identify multiple species in aerial imagery. It addresses the limitations of traditional, labor-intensive wildlife population assessments that are central to advancing environmental conservation effortsworldwide. While a number of methods exist to automate this process, they are often limited in their ability to generalize to different species or landscapes due to the dominance of homogeneous backgrounds and/or poorly captured local image structures. WM introduces two novel modules that help to capture the local structure and context of objects of interest to accurately localize and identify them, achieving a state-of-the-art (SOTA) detection rate of 0.56 mAP. Further, we introduce a large aerial imagery dataset with more than 11k Images and 28k annotations verified by trained experts. WM also achieves SOTA performance on 3 other publicly available aerial survey datasets collected across 4 different countries, improving mAP by 42%. Source code and trained models are available at Github.

Video stabilization is a longstanding computer vision problem, particularly pixel-level synthesis solutions for video stabilization which synthesize full frames add to the complexity of this task. These techniques aim to stabilize videos by synthesizing full frames while enhancing the stability of the considered video. This intensifies the complexity of the task due to the distinct mix of unique motion profiles and visual content present in each video sequence, making robust generalization with fixed parameters difficult. In our study, we introduce a novel approach to enhance the performance of pixel-level synthesis solutions for video stabilization by adapting these models to individual input video sequences. The proposed adaptation exploits low-level visual cues accessible during test-time to improve both the stability and quality of resulting videos. We highlight the efficacy of our methodology of "test-time adaptation" through simple fine-tuning of one of these models, followed by significant stability gain via the integration of meta-learning techniques. Notably, significant improvement is achieved with only a single adaptation step. The versatility of the proposed algorithm is demonstrated by consistently improving the performance of various pixel-level synthesis models for video stabilization in real-world scenarios.

Data-Free Knowledge Distillation (DFKD) is a promising task to train high-performance small models to enhance actual deployment without relying on the original training data.Existing methods commonly avoid relying on private data by utilizing synthetic or sampled data.However, a long-overlooked issue is that the severe distribution shifts between their substitution and original data, which manifests as huge differences in the quality of images and class proportions.The harmful shifts are essentially the confounder that significantly causes performance bottlenecks.To tackle the issue, this paper proposes a novel perspective with causal inference to disentangle the student models from the impact of such shifts.By designing a customized causal graph, we first reveal the causalities among the variables in the DFKD task.Subsequently, we propose a Knowledge Distillation Causal Intervention (KDCI) framework based on the backdoor adjustment to de-confound the confounder.KDCI can be flexibly combined with most existing state-of-the-art baselines. Experiments in combination with six representative DFKD methods demonstrate the effectiveness of our KDCI, which can obviously help existing methods under almost all settings, e.g., improving the baseline by up to 15.54\% accuracy on the CIFAR-100 dataset.

Previous advances in vehicle re-identification (ReID) are mostly reported under favorable lighting conditions, while cross-day-and-night performance is neglected, which greatly hinders the development of related traffic intelligence applications. This work instead develops a novel Day-Night Dual-domain Modulation (DNDM) vehicle re-identification framework for day-night cross-domain traffic scenarios. Specifically, a unique night-domain glare suppression module is provided to attenuate the headlight glare from raw nighttime vehicle images. To enhance vehicle features under low-light environments, we propose a dual-domain structure enhancement module in the feature extractor, which enhances geometric structures between appearance features. To alleviate day-night domain discrepancies, we develop a cross-domain class awareness module that facilitates the interaction between appearance and structure features in both domains. In this work, we address the Day-Night cross-domain ReID (DN-ReID) problem and provide a new cross-domain dataset named DN-Wild, including day and night images of 2,286 identities, giving in total 85,945 daytime images and 54,952 nighttime images. Furthermore, we also take into account the matter of balance between day and night samples, and provide a dataset called DN-348. Exhaustive experiments demonstrate the robustness of the proposed framework in the DN-ReID problem. The code and benchmark will be released soon.

Object inpainting is a task that involves adding objects to real images and seamlessly compositing them. With the recent commercialization of products like Stable Diffusion and Generative Fill, inserting objects into images by using prompts has achieved impressive visual results. In this paper, we propose a prompt suggestion model to simplify the process of prompt input. When the user provides an image and a mask, our model predicts suitable prompts based on the partial contextual information in the masked image, and the shape and location of the mask. Specifically, we introduce a concept-diffusion in the CLIP space that predicts CLIP-text embeddings from a masked image. These diffused embeddings can be directly injected into open-source inpainting models like Stable Diffusion and its variants. Alternatively, they can be decoded into natural language for use in other publicly available applications such as Generative Fill. Our prompt suggestion model demonstrates a balanced accuracy and diversity, showing its capability to be both contextually aware and creatively adaptive.

The burgeoning field of Multimodal Large Language Models (MLLMs) has exhibited remarkable performance in diverse tasks such as captioning, commonsense reasoning, and visual scene understanding. However, the deployment of these large-scale MLLMs on client devices is hindered by their extensive model parameters, leading to a notable decline in generalization capabilities when these models are compressed for device deployment. Addressing this challenge, we introduce a Cloud-Device Collaborative Continual Adaptation framework, designed to enhance the performance of compressed, device-deployed MLLMs by leveraging the robust capabilities of cloud-based, larger-scale MLLMs.Our framework is structured into three key components: a device-to-cloud uplink for efficient data transmission, cloud-based knowledge adaptation, and an optimized cloud-to-device downlink for model deployment. In the uplink phase, we employ an Uncertainty-guided Token Sampling (UTS) strategy to effectively filter out-of-distribution tokens, thereby reducing transmission costs and improving training efficiency. On the cloud side, we propose Adapter-based Knowledge Distillation (AKD) method to transfer refined knowledge from large-scale to compressed, pocket-size MLLMs. Furthermore, we propose a Dynamic Weight update Compression (DWC) strategy for the downlink, which adaptively selects and quantizes updated weight parameters, enhancing transmission efficiency and reducing the representational disparity between cloud and device models. Extensive experiments on several multimodal benchmarks demonstrate the superiority of our proposed framework over prior Knowledge Distillation and device-cloud collaboration methods. Notably, we also validate the feasibility of our approach to real-world experiments.

Visual perception evolves over time. This is particularly the case of oracle bone scripts, where visual glyphs seem intuitive to people from distant past prove difficult to be understood in contemporary eyes. While semantic correspondence of an oracle can be found via a dictionary lookup, this proves to be not enough for public viewers to connect the dots, i.e., why does this oracle mean that? Common solution relies on a laborious curation process to collect visual guide for each oracle (Fig.1), which hinges on the case-by-case effort and taste of curators. This paper delves into one natural follow-up question: can AI take over? Begin with a comprehensive human study, we show participants could indeed make better sense of an oracle glyph subjected to a proper visual guide and its efficacy can be approximated via a novel metric termed TransOV (Transferable Oracle Visuals). We then define a new conditional visual generation task based on an oracle glyph and its semantic meaning and importantly approach it by circumventing any form of model training in the presence of fatal lack of oracle data. At its heart is to leverage foundation model like GPT-4V to reason about the visual cues hidden inside an oracle and take advantage of an existing text-to-image model for final visual guide generation. Extensive empirical evidence shows our AI-enabled visual guides achieve significantly comparable TransOV performance compared with those collected under manual efforts. Finally, we demonstrate the versatility of our system under a more complex setting, where it is required to work alongside an AI image denoiser to cope with raw oracle scan image inputs (cf. processed clean oracle glyphs). The code, data, and model will be made publicly available.

Detecting objects in low-light scenarios presents a persistent challenge, as detectors trained on well-lit data exhibit significant performance degradation on low-light data due to low visibility. Previous methods mitigate this issue by exploring image enhancement or object detection techniques with real low-light image datasets. However, the progress is impeded by the inherent difficulties about collecting and annotating low-light images. To address this challenge, we propose to boost low-light object detection with zero-shot day-night domain adaptation, which aims to generalize a detector from well-lit scenarios to low-light ones without requiring real low-light data. Revisiting Retinex theory in the low-level vision, we first design a reflectance representation learning module to learn Retinex-based illumination invariance in images with a carefully designed illumination invariance reinforcement strategy. Next, an interchange-redecomposition-coherence procedure is introduced to improve over the vanilla Retinex image decomposition process by performing two sequential image decompositions and introducing a redecomposition cohering loss. Extensive experiments on ExDark, DARK FACE, and CODaN datasets show strong low-light generalizability of our method. Our code is available at https://github.com/ZPDu/DAI-Net.

We propose InNeRF360, an automatic system that accurately removes text-specified objects from 360-degree Neural Radiance Fields (NeRF). The challenge is to effectively remove objects while inpainting perceptually consistent content for the missing regions, which is particularly demanding for existing NeRF models due to their implicit volumetric representation. Moreover, unbounded scenes are more prone to floater artifacts in the inpainted region than frontal-facing scenes, as the change of object appearance and background across views is more sensitive to inaccurate segmentations and inconsistent inpainting. With a trained NeRF and a text description, our method efficiently removes specified objects and inpaints visually consistent content without artifacts. We apply depth-space warping to enforce consistency across multiview text-encoded segmentations, and then refine the inpainted NeRF model using perceptual priors and 3D diffusion-based geometric priors to ensure visual plausibility. Through extensive experiments in segmentation and inpainting on 360-degree and frontal-facing NeRFs, we show that InNeRF360 is effective and enhances NeRF's editability. Project page: https://ivrl.github.io/InNeRF360.

Vision-language models (VLMs) pre-trained on web-scale datasets have demonstrated remarkable capabilities on downstream tasks when fine-tuned with minimal data. However, many VLMs rely on proprietary data and are not open-source, which restricts the use of white-box approaches for fine-tuning. As such, we aim to develop a black-box approach to optimize VLMs through natural language prompts, thereby avoiding the need to access model parameters, feature embeddings, or even output logits. We propose employing chat-based LLMs to search for the best text prompt for VLMs. Specifically, we adopt an automatic "hill-climbing" procedure that converges to an effective prompt by evaluating the performance of current prompts and asking LLMs to refine them based on textual feedback, all within a conversational process without human-in-the-loop. In a challenging 1-shot image classification setup, our simple approach surpasses the white-box continuous prompting method (CoOp) by an average of 1.5% across 11 datasets including ImageNet. Our approach also outperforms both human-engineered and LLM-generated prompts. We highlight the advantage of conversational feedback that incorporates both positive and negative prompts, suggesting that LLMs can utilize the implicit "gradient" direction in textual feedback for a more efficient search. In addition, we find that the text prompts generated through our strategy are not only more interpretable but also transfer well across different VLM architectures in a black-box manner. Lastly, we demonstrate our framework on a state-of-the-art black-box VLM (DALLE-3) for text-to-image optimization.

Crack segmentation datasets make great efforts to obtain the ground truth crack or non-crack labels as clearly as possible. However, it can be observed that ambiguities are still inevitable when considering the marginal non-crack region, due to low contrast and heterogeneous texture. To solve this problem, we propose a novel clustering-inspired representation learning framework, which contains a two-phase strategy for automatic crack segmentation. In the first phase, a pre-process is proposed to localize the marginal non-crack region. Then, we propose an ambiguity-aware segmentation loss (Aseg Loss) that enables crack segmentation models to capture ambiguities in the above regions via learning segmentation variance, which allows us to further localize ambiguous regions. In the second phase, to learn the discriminative features of the above regions, we propose a clustering-inspired loss (CI Loss) that alters the supervision learning of these regions into an unsupervised clustering manner. We demonstrate that the proposed method could surpass the existing crack segmentation models on various datasets and our constructed CrackSeg5k dataset.

We present InstructDiffusion, a unified and generic framework for aligning computer vision tasks with human instructions. Unlike existing approaches that integrate prior knowledge and pre-define the output space (e.g., categories and coordinates) for each vision task, we cast diverse vision tasks into a human-intuitive image-manipulating process whose output space is a flexible and interactive pixel space. Concretely, the model is built upon the diffusion process and is trained to predict pixels according to user instructions, such as encircling the man's left shoulder in red or applying a blue mask to the left car. InstructDiffusion could handle a variety of vision tasks, including understanding tasks (such as segmentation and keypoint detection) and generative tasks (such as editing and enhancement) and outperforms prior methods on novel datasets. This represents a solid step towards a generalist modeling interface for vision tasks, advancing artificial general intelligence in the field of computer vision.

Templates serve as a good starting point to implement a design (e.g., banner, slide) but it takes great effort from designers to manually create. In this paper, we present Desigen, an automatic template creation pipeline which generates background images as well as harmonious layout elements over the background. Different from natural images, a background image should preserve enough non-salient space for the overlaying layout elements. To equip existing advanced diffusion-based models with stronger spatial control, we propose two simple but effective techniques to constrain the saliency distribution and reduce the attention weight in desired regions during the background generation process. Then conditioned on the background, we synthesize the layout with a Transformer-based autoregressive generator. To achieve a more harmonious composition, we propose an iterative inference strategy to adjust the synthesized background and layout in multiple rounds. We construct a design dataset with more than 40k advertisement banners to verify our approach. Extensive experiments demonstrate that the proposed pipeline generates high-quality templates comparable to human designers. More than a single-page design, we further show an application of presentation generation that outputs a set of theme-consistent slides. The data and code will be released.

Backdoor attacks have been well-studied in visible light object detection (VLOD) in recent years. However, VLOD can not effectively work in dark and temperature-sensitive scenarios. Instead, thermal infrared object detection (TIOD) is the most accessible and practical in such environments. In this paper, our team is the first to investigate the security vulnerabilities associated with TIOD in the context of backdoor attacks, spanning both the digital and physical realms. We introduce two novel types of backdoor attacks on TIOD, each offering unique capabilities: Object-affecting Attack and Range-affecting Attack. We conduct a comprehensive analysis of key factors influencing trigger design, which include temperature, size, material, and concealment. These factors, especially temperature, significantly impact the efficacy of backdoor attacks on TIOD. A thorough understanding of these factors will serve as a foundation for designing physical triggers and temperature controlling experiments. Our study includes extensive experiments conducted in both digital and physical environments. In the digital realm, we evaluate our approach using benchmark datasets for TIOD, achieving an Attack Success Rate (ASR) of up to 98.21%. In the physical realm, we test our approach in two real-world settings: a traffic intersection and a parking lot, using a thermal infrared camera. Here, we attain an ASR of up to 98.38%.

In this paper, we present a novel indoor 3D reconstruction method with occluded surface completion, given a sequence of depth readings. Prior state-of-the-art (SOTA) methods only focus on the reconstruction of the visible areas in a scene, neglecting the invisible areas due to the occlusions, e.g., the contact surface between furniture, occluded wall and floor. Our method tackles the task of completing the occluded scene surfaces, resulting in a complete 3D scene mesh. The core idea of our method is learning 3D geometry prior from various complete scenes to infer the occluded geometry of an unseen scene from solely depth measurements. We design a coarse-fine hierarchical octree representation coupled with a dual-decoder architecture, i.e., Geo-decoder and 3D Inpainter, which jointly reconstructs the complete 3D scene geometry. The Geo-decoder with detailed representation at fine levels is optimized online for each scene to reconstruct visible surfaces. The 3D Inpainter with abstract representation at coarse levels is trained offline using various scenes to complete occluded surfaces. As a result, while the Geo-decoder is specialized for an individual scene, the 3D Inpainter can be generally applied across different scenes. We evaluate the proposed method on the 3D Completed Room Scene (3D-CRS) and iTHOR datasets, significantly outperforming the SOTA methods by a gain of 16.8% and 24.2% in terms of the completeness of 3D reconstruction.

Astronaut photography, spanning six decades of human spaceflight, presents a unique Earth observations dataset with immense value for both scientific research and disaster response. Despite its significance, accurately localizing the geographical extent of these images, crucial for effective utilization, poses substantial challenges. Current manual localization efforts are time-consuming, motivating the need for automated solutions. We propose a novel approach – leveraging image retrieval – to address this challenge efficiently. We introduce innovative training techniques, including Year-Wise Data Augmentation and a Neutral-Aware Multi-Similarity Loss, which contribute to the development of a high-performance model, EarthLoc. We develop six evaluation datasets and perform a comprehensive benchmark comparing EarthLoc to existing methods, showcasing its superior efficiency and accuracy. Our approach marks a significant advancement in automating the localization of astronaut photography, which will help bridge a critical gap in Earth observations data.

As manipulating images may lead to misinterpretation of the visual content, addressing the image forgery detection and localization (IFDL) problem has drawn serious public concerns. In this work, we propose a simple assumption that the effective forensic method should focus on the mesoscopic properties of images. Based on the assumption, a novel two-stage self-supervised framework leveraging the diffusion model for IFDL task, \ie, DiffForensics, is proposed in this paper. The DiffForensics begins with self-supervised denoising diffusion paradigm equipped with the module of encoder-decoder structure, by freezing the pre-trained encoder (\eg, in ADE-20K) to inherit macroscopic features for general image characteristics, while encouraging the decoder to learn microscopic feature representation of images, enforcing the whole model to focus the mesoscopic representations. The pre-trained model as a prior, is then further fine-tuned for IFDL task with the customized Edge Cue Enhancement Module (ECEM), which progressively highlights the boundary features within the manipulated regions, thereby refining tampered area localization with better precision. Extensive experiments on several public challenging datasets demonstrate the effectiveness of the proposed method compared with other state-of-the-art methods. The proposed DiffForensics could significantly improve the model’s capabilities for both accurate tamper detection and precise tamper localization while concurrently elevating its generalization and robustness.

Music recommendation for videos attracts growing interest in multi-modal research. However, existing systems focus primarily on content compatibility, often ignoring the users' preferences. Their inability to interact with users for further refinements or to provide explanations leads to a less satisfying experience. We address these issues with MuseChat, a first-of-its-kind dialogue-based recommendation system that personalizes music suggestions for videos. Our system consists of two key functionalities with associated modules: recommendation and reasoning. The recommendation module takes a video along with optional information including previous suggested music and user's preference as inputs and retrieves an appropriate music matching the context. The reasoning module, equipped with the power of Large Language Model (Vicuna-7B) and extended to multi-modal inputs, is able to provide reasonable explanation for the recommended music. To evaluate the effectiveness of MuseChat, we build a large-scale dataset, conversational music recommendation for videos, that simulates a two-turn interaction between a user and a recommender based on accurate music track information. Experiment results show that MuseChat achieves significant improvements over existing video-based music retrieval methods as well as offers strong interpretability and interactability. The dataset of this work is available at https://dongzhikang.github.io/musechat.

Pose refinement is an interesting and practically relevant research direction. Pose refinement can be used to (1) obtain a more accurate pose estimate from an initial prior (e.g., from retrieval), (2) as pre-processing, i.e., to provide a better starting point to a more expensive pose estimator, (3) as post-processing of a more accurate localizer. Existing approaches focus on learning features / scene representations for the pose refinement task. This involves training an implicit scene representation or learning features while optimizing a camera pose-based loss. A natural question is whether training specific features / representations is truly necessary or whether similar results can be already achieved with more generic features. In this work, we present a simple approach that combines pre-trained features with a particle filter and a renderable representation of the scene. Despite its simplicity, it achieves state-of-the-art results, demonstrating that one can easily build a pose refiner without the need for specific training. The code will be released upon acceptance.

A novel approach to blind image quality assessment, called quality comparison network (QCN), is proposed in this paper, which sorts the feature vectors of input images according to their quality scores in an embedding space. QCN employs comparison transformers (CTs) and score pivots, which act as the centroids of feature vectors of similar-quality images. Each CT updates the score pivots and the feature vectors of input images based on their ordered correlation. To this end, we adopt four loss functions. Then, we estimate the quality score of a test image by searching the nearest score pivot to its feature vector in the embedding space. Extensive experiments show that the proposed QCN algorithm yields excellent image quality assessment performances on various datasets. Furthermore, QCN achieves great performances in cross-dataset evaluation, demonstrating its superb generalization capability. The source codes are available at https://github.com/nhshin-mcl/QCN.

Crowd counting is a fundamental problem in crowd analysis which is typically accomplished by estimating a crowd density map and summing over the density values. However, this approach suffers from background noise accumulation and loss of density due to the use of broad Gaussian kernels to create the ground truth density maps. This issue can be overcome by narrowing the Gaussian kernel. However, existing approaches perform poorly when trained with ground truth density maps with broad kernels. To deal with this limitation, we propose using conditional diffusion models to predict density maps, as diffusion models show high fidelity to training data during generation. With that, we present $CrowdDiff$ that generates the crowd density map as a reverse diffusion process. Furthermore, as the intermediate time steps of the diffusion process are noisy, we incorporate a regression branch for direct crowd estimation only during training to improve the feature learning. In addition, owing to the stochastic nature of the diffusion model, we introduce producing multiple density maps to improve the counting performance contrary to the existing crowd counting pipelines. We conduct extensive experiments on publicly available datasets to validate the effectiveness of our method. $CrowdDiff$ outperforms existing state-of-the-art crowd counting methods on several public crowd analysis benchmarks with significant improvements. Code and pre-trained models will be made publicly available after the review process.

In Federated Learning (FL), the data in each client is typically assumed fixed or static. However, data often comes in an incremental manner in real-world applications, where the data domain may increase dynamically. In this work, we study catastrophic forgetting with data heterogeneity in Federated Incremental Learning (FIL) scenarios where edge clients may lack enough storage space to retain full data. We propose to employ a simple, generic framework for FIL named Re-Fed, which can coordinate each client to cache important samples for replay. More specifically, when a new task arrives, each client first caches selected previous samples based on their global and local importance. Then, the client trains the local model with both the cached samples and the samples from the new task. Theoretically, we analyze the ability of Re-Fed to discover important samples for replay thus alleviating the catastrophic forgetting problem. Moreover, we empirically show that Re-Fed achieves competitive performance compared to state-of-the-art methods.

Limited training data is a long-standing problem for video emotion analysis (VEA). Existing works leverage the power of large-scale image datasets for transferring while failing to extract the temporal correlation of affective cues in the video. Inspired by psychology research and empirical theory, we verify that the degree of emotion may vary in different segments of the video, thus introducing the sentiment complementary and emotion intrinsic among temporal segments. We propose an MAE-style method for learning robust affective representation of videos via masking, termed MART. First, we extract the affective cues of the lexicon and verify the extracted one by computing its matching score with video content, in terms of sentiment and emotion scores alongside the temporal dimension. Then, with the verified cues, we propose masked affective modeling to recover temporal emotion distribution. We present temporal affective complementary learning that pulls the complementary part and pushes the intrinsic one of masked multimodal features, where the constraint is set with cross-modal attention among features to mask the video and recover the degree of emotion among segments. Extensive experiments on five benchmarks show the superiority of our method in video sentiment analysis, video emotion recognition, multimodal sentiment analysis, and multimodal emotion recognition.

In radio astronomy, visibility data, which are measurements of wave signals from radio telescopes, are transformed into images for observation of distant celestial objects. However, these resultant images usually contain both real sources and artifacts, due to signal sparsity and other factors. One way to obtain cleaner images is to reconstruct samples into dense forms before imaging. Unfortunately, existing reconstruction methods often miss some components of visibility in frequency domain, so blurred object edges and persistent artifacts remain in the images. Furthermore, the computation overhead is high on irregular visibility samples due to the data skew. To address these problems, we propose PolarRec, a transformer-encoder-conditioned reconstruction pipeline with visibility samples converted into the polar coordinate system. This coordinate system matches the way in which radio telescopes observe a celestial area as the Earth rotates. As a result, visibility samples distribute in the polar system more uniformly than in the Cartesian space. Therefore, we propose to use radial distance in the loss function, to help reconstruct complete visibility effectively. Also, we group visibility samples by their polar angles and propose a group-based encoding scheme to improve the efficiency. Our experiments demonstrate that PolarRec markedly improves imaging results by faithfully reconstructing all frequency components in the visibility domain while significantly reducing the computation cost in visibility data encoding. The code is available at https://github.com/RapidsAtHKUST/PolarRec.

This paper addresses the challenge of object-centric layout generation under spatial constraints, seen in multiple domains including floorplan design process. The design process typically involves specifying a set of spatial constraints that include object attributes like size and inter-object relations such as relative positioning. Existing works, which typically represent objects as single nodes, lack the granularity to accurately model complex interactions between objects. For instance, often only certain parts of an object, like a room's right wall, interact with adjacent objects. To address this gap, we introduce a factor graph based approach with four latent variable nodes for each room, and a factor node for each constraint. The factor nodes represent dependencies among the variables to which they are connected, effectively capturing constraints that are potentially of a higher order. We then develop message-passing on the bipartite graph, forming a factor graph neural network that is trained to produce a floorplan that aligns with the desired requirements. Our approach is simple and generates layouts faithful to the user requirements, demonstrated by a large improvement in IOU scores over existing methods. Additionally, our approach, being inferential and accurate, is well-suited to the practical human-in-the-loop design process where specifications evolve iteratively, offering a practical and powerful tool for AI-guided design.

In-context prompting in large language models (LLMs) has become a prevalent approach to improving zero-shot capabilities, but this idea is less explored in vision domain. Existing visual prompting methods focus on referring segmentation to segment the most relevant object, falling short of addressing many general vision tasks like open-set segmentation and detection. In this paper, we introduce a unified visual in-context prompting framework for both tasks, as shown in Fig. 1. In particular, we build on top of an encoder-decoder architecture, and develop a versatile content embedder to support a variety of prompts like strokes, boxes, and points. We further enhance it to be able to take an arbitrary number of reference images as the context. Our explorations show that the proposed in-context prompting demonstrates impressive referring and generic segmentation capabilities to refer and detect, yielding competitive performance to close-set in-domain datasets and showing promising results on many open-set segmentation datasets.

Federated continual learning (FCL) is a typical mechanism to achieve collaborative model training among clients that own dynamic data. While traditional FCL methods have been proved effective, they do not consider the task repeatability and fail to achieve good performance under this practical scenario. In this paper, we propose a new paradigm, namely \textit{Traceable Federated Continual Learning (TFCL)}, aiming to cope with repetitive tasks by tracing and augmenting them. Following the new paradigm, we develop \textbf{TagFed}, a framework that enables accurate and effective \textbf{T}racing, \textbf{a}u\textbf{g}mentation, and \textbf{Fed}eration for TFCL. The key idea is to decompose the whole model into a series of marked sub-models for optimizing each client task, before conducting group-wise knowledge aggregation, such that the repetitive tasks can be located precisely and federated selectively for improved performance. Extensive experiments on our constructed benchmark demonstrate the effectiveness and efficiency of the proposed framework \footnote{ Source code will be released after notification.}.

Advanced life forms, sustained by the synergistic interaction of neural cognitive mechanisms, continually acquire and transfer knowledge throughout their lifespan. In contrast, contemporary machine learning paradigms exhibit limitations in emulating the facets of continual learning (CL). Nonetheless, the emergence of large language models (LLMs) presents promising avenues for realizing CL via interactions with these models.Drawing on Complementary Learning System theory, this paper presents a novel Interactive Continual Learning (ICL) framework, enabled by collaborative interactions among models of various sizes. Specifically, we assign the ViT model as System1 and multimodal LLM as System2.To enable the memory module to deduce tasks from class information and enhance Set2Set retrieval, we propose the Class-Knowledge-Task Multi-Head Attention (CKT-MHA).Additionally, to improve memory retrieval in System1 through enhanced geometric representation, we introduce the CL-vMF mechanism, based on the von Mises-Fisher (vMF) distribution. Meanwhile, we introduce the von Mises-Fisher Outlier Detection and Interaction (vMF-ODI) strategy to identify hard examples, thus enhancing collaboration between System1 and System2 for complex reasoning realization.Comprehensive evaluation of our proposed ICL demonstrates significant resistance to forgetting and superior performance relative to existing methods.

Planet-scale image geolocalization remains a challenging problem due to the diversity of images originating from anywhere in the world. Although approaches based on vision transformers have made significant progress in geolocalization accuracy, success in prior literature is constrained to narrow distributions of images of landmarks, and performance has not generalized to unseen places. We present a new geolocalization system that combines semantic geocell creation, multi-task contrastive pretraining, and a novel loss function. Additionally, our work is the first to perform retrieval over location clusters for guess refinements. We train two models for evaluations on street-level data and general-purpose image geolocalization; the first model, PIGEON, is trained on data from the game of GeoGuessr and is capable of placing over 40\% of its guesses within 25 kilometers of the target location globally. We also develop a bot and deploy PIGEON in a blind experiment against humans, ranking in the top 0.01\% of players. We further challenge one of the world's foremost professional GeoGuessr players to a series of six matches with millions of viewers, winning all six games. Our second model, PIGEOTTO, differs in that it is trained on a dataset of images from Flickr and Wikipedia, achieving state-of-the-art results on a wide range of image geolocalization benchmarks, outperforming the previous SOTA by up to 7.7 percentage points on the city accuracy level and up to 38.8 percentage points on the country level. Our findings suggest that PIGEOTTO is the first image geolocalization model that effectively generalizes to unseen places and that our approach can pave the way for highly accurate, planet-scale image geolocalization systems. Our code is available on GitHub.

Referring Image Segmentation~(RIS) aims to segment an object described by a language expression from an image. Recently, state-of-art Transformer-based methods have been proposed to efficiently leverage cross-modal dependencies, enhancing performance for referring segmentation. Specifically, all these transformer based methods predict masks, where each query learn different objects. However, as the prediction is single-mask, it leads to Query collapse, where all query leads to same mask prediction. To address these limitations, we propose a Multi-modal Query Feature Fusion technique with two key designs: (1) Gaussian enhanced Multi-modal Fusion, a novel visual grounding mechanism for extracting rich local visual information and modeling global visual linguistic relationships in an integrated manner. (2) Language-Query Selection Module for generating diverse set of queries and scoring network that selectively updates only queries expected to be referenced by decoder. In addition, we also show that adding an auxiliary loss, to increase the distance between mask representation of Queries, help improving the performance. Extensive experiments on four benchmark datasets demonstrate the effectiveness of our framework.

While existing large vision-language multimodal models focus on whole image understanding, there is a prominent gap in achieving region-specific comprehension. Current approaches that use textual coordinates or spatial encodings often fail to provide a user-friendly interface for visual prompting. To address this challenge, we introduce a novel multimodal model capable of decoding arbitrary (free-form) visual prompts. This allows users to intuitively mark images and interact with the model using natural cues like a red bounding box or pointed arrow. Our simple design directly overlays visual markers onto the RGB image, eliminating the need for complex region encodings, yet achieves state-of-the-art performance on region-understanding tasks like Visual7W, PointQA, and Visual Commonsense Reasoning benchmark. Furthermore, we present RegionBench, a comprehensive benchmark to assess the capability of models in understanding visual prompts across multiple dimensions, enabling future research in this domain. Code and demo will be released.

This work breaks through the Base-New Tradeoff (BNT)dilemma in prompt tuning, i.e., the better the tuned model generalizes to the base (or target) task, the worse it generalizes to new tasks, and vice versa. Specifically, through an in-depth analysis of the learned features of the base and new tasks, we observe that the BNT stems from a channel bias issue, i.e., the vast majority of feature channels are occupied by base-specific knowledge, resulting in the collapse of taskshared knowledge important to new tasks. To address this, we propose the Decoupled Prompt Tuning (DePT) framework, which decouples base-specific knowledge from feature channels into an isolated feature space during prompt tuning, so as to maximally preserve task-shared knowledge in the original feature space for achieving better zero-shot generalization on new tasks. Notably, our DePT is orthogonal to existing prompt tuning methods, and can enhance them with negligible additional computational cost. Extensive experiments on 11 diverse datasets show the strong flexibility and effectiveness of DePT.Code: https://anonymous.4open.science/r/DePT.

Existing approaches to video understanding, mainly designed for short videos from a third-person perspective, are limited in their applicability in certain fields, such as robotics. In this paper, we delve into open-ended question-answering (QA) in long, egocentric videos, which allows individuals or robots to inquire about their own past visual experiences. This task presents unique challenges, including the complexity of temporally grounding queries within extensive video content, the high resource demands for precise data annotation, and the inherent difficulty of evaluating open-ended answers due to their ambiguous nature. Our proposed approach tackles these challenges by (i) integrating query grounding and answering within a unified model to reduce error propagation; (ii) employing large language models for efficient and scalable data synthesis; and (iii) introducing a close-ended QA task for evaluation, to manage answer ambiguity. Extensive experiments demonstrate the effectiveness of our method, which also achieves state-of-the-art performance on the QAEgo4D and Ego4D-NLQ benchmarks. Code, data, and models are open-sourced at https://github.com/Becomebright/GroundVQA.

Multi-modal Large Language Models (MLLMs) tuned on machine-generated instruction-following data have demonstrated remarkable performance in various multi-modal understanding and generation tasks. However, the hallucinations inherent in machine-generated data, which could lead to hallucinatory outputs in MLLMs, remain under-explored. This work aims to investigate various hallucinations (i.e., object, relation, attribute hallucinations) and mitigate those hallucinatory toxicities in large-scale machine-generated visual instruction datasets. Drawing on the human ability to identify factual errors, we present a novel hallucination detection and elimination framework, HalluciDoctor, based on the cross-checking paradigm. We use our framework to identify and eliminate hallucinations in the training data automatically. Interestingly, HalluciDoctor also indicates that spurious correlations arising from long-tail object co-occurrences contribute to hallucinations. Based on that, we execute counterfactual visual instruction expansion to balance data distribution, thereby enhancing MLLMs' resistance to hallucinations. Comprehensive experiments on hallucination evaluation benchmarks show that our method successfully mitigates 44.6% hallucinations relatively and maintains competitive performance compared to LLaVA. The data and code for this paper are publicly available.

Recent breakthroughs in vision-language models (VLMs) start a new page in the vision community. The VLMs provide stronger and more generalizable feature embeddings compared to those from ImageNet-pretrained models, thanks to the training on the large-scale Internet image-text pairs. However, despite the amazing achievement from the VLMs, vanilla Vision Transformers (ViTs) remain the default choice for the image encoder. Although pure transformer proves its effectiveness in the text encoding area, it remains questionable whether it is also the case for image encoding, especially considering that various types of networks are proposed on the ImageNet benchmark, which, unfortunately, are rarely studied in VLMs. Due to small data/model scale, the original conclusions of model design on ImageNet can be limited and biased. In this paper, we aim at building an evaluation protocol of vision models in the vision-language era under the contrastive language-image pretraining (CLIP) framework. We provide a comprehensive way to benchmark different vision models, covering their zero-shot performance and scalability in both model and training data sizes. To this end, we introduce ViTamin, a new vision models tailored for VLMs. ViTamin-L significantly outperforms ViT-L by 2.0% ImageNet zero-shot accuracy, when using the same publicly available DataComp-1B dataset and the same OpenCLIP training scheme. ViTamin-L presents promising results on 60 diverse benchmarks, including classification, retrieval, open-vocabulary detection and segmentation, and large multi-modal models. When further scaling up the model size, our ViTamin-XL with only 436M parameters attains 82.9% ImageNet zero-shot accuracy, surpassing 82.0% achieved by EVA-E that has ten times more parameters (4.4B).

In the past few decades, Japanese comics, commonly referred to as Manga, have transcended both cultural and linguistic boundaries to become a true worldwide sensation. Yet, the inherent reliance on visual cues and illustration within manga renders it largely inaccessible to individuals with visual impairments. In this work, we seek to address this substantial barrier, with the aim of ensuring that manga can be appreciated and actively engaged by everyone. Specifically, we tackle the problem of diarisation i.e. generating a transcription of who said what and when, in a fully automatic way.To this end, we make the following contributions: (1) we present a unified model, Magi, that is able to (a) detect panels, text boxes and character boxes, (b) cluster characters by identity (without knowing the number of clusters apriori), and (c) associate dialogues to their speakers; (2) we propose a novel approach that is able to sort the detected text boxes in their reading order and generate a dialogue transcript; (3) we annotate an evaluation benchmark for this task using publicly available [English] manga pages.

Integration of Large Language Models (LLMs) into visual domain tasks, resulting in visual-LLMs (V-LLMs), has enabled exceptional performance in vision-language tasks, particularly for visual question answering (VQA). However, existing V-LLMs (e.g. BLIP-2, LLaVA) demonstrate weak spatial reasoning and localization awareness. Despite generating highly descriptive and elaborate textual answers, these models fail at simple tasks like distinguishing a left vs right location. In this work, we explore how image-space coordinate based instruction fine-tuning objectives could inject spatial awareness into V-LLMs. We discover optimal coordinate representations, data-efficient instruction fine-tuning objectives, and pseudo-data generation strategies that lead to improved spatial awareness in V-LLMs. Additionally, our resulting model improves VQA across image and video domains, reduces undesired hallucination, and generates better contextual object descriptions. Experiments across 5 vision-language tasks involving 14 different datasets establish the clear performance improvements achieved by our proposed framework.

Vision-language models (VLMs) such as CLIP excel in zero-shot recognition but exhibit drastically imbalanced performance across visual concepts in downstream tasks. For example, despite an impressive mean zero-shot accuracy on ImageNet (72.7\%), CLIP yields $<$10\% on ten concepts (e.g., ${\tt night}$ ${\tt snake}$ and ${\tt snoek}$). This is presumably because these concepts are under-represented in VLMs' pretraining datasets, which are believed to exhibit imbalanced distributions of concepts. Yet, assessing this imbalance is challenging, as calculating the frequency of specific concepts within large-scale pretraining data is not straightforward. In this work, we make the first attempt to measure the concept frequency in VLMs' pretraining data by counting relevant pretraining texts. We also use off-the-shelf language models to help count relevant texts that contain synonyms of the given concepts and resolve ambiguous cases. Our analysis confirms that visual concepts follow a long-tailed distribution in the pretraining data, which strongly correlates with per-class accuracies. Further, to mitigate VLMs' imbalanced performance in zero-shot recognition, we propose ${\bf RE}$trieval-${\bf A}$ugmented ${\bf L}$earning (REAL). First, instead of prompting VLMs using the original class names defined in a downstream task, REAL uses their most frequent synonyms found in the pretraining texts. This already outperforms human-engineered and LLM-generated prompts over nine benchmark datasets, likely because VLMs have seen more images associated with the more frequent synonyms. Second, REAL uses all the concept synonyms to retrieve a small class-balanced subset of images from the pretraining data to learn a robust classifier. REAL rivals the recent retrieval-augmented solution REACT, using $400\times$ less storage and 10,000$\times$ less training time!

Referring expression segmentation (RES) aims at segmenting the foreground masks of the entities that match the descriptive natural language expression. Previous datasets and methods for classic RES task heavily rely on the prior assumption that one expression must refer to object-level targets. In this paper, we take a step further to finer-grained part-level RES task. To promote the object-level RES task towards finer-grained vision-language understanding, we put forward a new multi-granularity referring expression segmentation (MRES) task and construct an evaluation benchmark called RefCOCOm by manual annotations. By employing our automatic model-assisted data engine, we build the largest visual grounding dataset namely MRES-32M, which comprises over 32.2M high-quality masks and captions on the provided 1M images. Besides, a simple yet strong model named UniRES is designed to accomplish the unified object-level and part-level grounding task. Extensive experiments on our RefCOCOm for MRES and three datasets (i.e., RefCOCO(+/g)) for classic RES task demonstrate the superiority of our method over previous state-of-the-art methods. To foster future research into fine-grained visual grounding, our benchmark RefCOCOm, the MRES-32M dataset and model UniRES will be publicly available.

Large Multimodal Models (LMMs) extend Large Language Models to the vision domain. Initial LMMs used holistic images and text prompts to generate ungrounded textual responses. Very recently, region-level LMMs have been used to generate visually grounded responses. However, they are limited to only referring to a single object category at a time, require users to specify the regions in inputs, or cannot offer dense pixel-wise object grounding. In this work, we present Grounding LMM (GLaMM), the first model that can generate natural language responses seamlessly intertwined with corresponding object segmentation masks. GLaMM not only grounds objects appearing in the conversations but is flexible enough to accept both textual and optional visual prompts (region of interest) as input. This empowers users to interact with the model at various levels of granularity, both in textual and visual domains. Due to the lack of standard benchmarks for the novel setting of visually Grounded Conversation Generation (GCG), we introduce a comprehensive evaluation protocol with our curated grounded conversations. Our proposed GCG task requires densely grounded concepts in natural scenes at a large-scale. To this end, we propose a densely annotated Grounding-anything Dataset (GranD) using our proposed automated annotation pipeline that encompasses 7.5M unique concepts grounded in a total of 810M regions available with segmentation masks. Besides GCG, GLaMM also performs effectively on several downstream tasks, e.g., referring expression segmentation, image and region-level captioning and vision-language conversations. Our codes, data and models will be publicly released.

Contrastive Language-Image Pre-training (CLIP) plays an essential role in extracting valuable content information from images across diverse tasks. It aligns textual and visual modalities to comprehend the entire image, including all the details, even those irrelevant to specific tasks. However, for a finer understanding and controlled editing of images, it becomes crucial to focus on specific regions of interest, which can be indicated as points, masks, or boxes by humans or perception models. To fulfill the requirements, we introduce Alpha-CLIP, an enhanced version of CLIP with an auxiliary alpha channel to suggest attentive regions and fine-tuned with constructed millions of RGBA region-text pairs. Alpha-CLIP not only preserves the visual recognition ability of CLIP but also enables precise control over the emphasis of image contents. It demonstrates effectiveness in various tasks, including but not limited to open-world recognition, multimodal large language models, and conditional 2D / 3D generation. It has a strong potential to serve as a versatile tool for image-related tasks. All the code, models, and training data will be publicly available.

Large language models have achieved great success in recent years, so as their variants in vision. Existing vision-language models can describe images in natural languages, answer visual-related questions, or perform complex reasoning about the image. However, it is yet unclear how localization tasks, such as word grounding or referring localization, can be performed using large language models. In this work, we aim to develop a vision-language model that can take locations, for example, a set of points or boxes, as either inputs or outputs. When taking locations as inputs, the model performs location-conditioned captioning, which generates captions for the indicated object or region. When generating locations as outputs, our model regresses pixel coordinates for each output word generated by the language model, and thus performs dense word grounding. Our model is pre-trained on the Localized Narrative dataset, which contains pixel-word-aligned captioning from human attention. We show our model can be applied to various location-aware vision-language tasks, including referring localization, location-conditioned captioning, and dense object captioning, archiving state-of-the-art performance on RefCOCO and Visual Genome.

Multi-modal Large Language Models (MLLMs) have demonstrated impressive instruction abilities across various open-ended tasks. However, previous methods have primarily focused on enhancing multi-modal capabilities. In this work, we introduce a versatile multi-modal large language model, mPLUG-Owl2, which effectively leverages modality collaboration to improve performance in both text and multi-modal tasks. mPLUG-Owl2 utilizes a modularized network design, with the language decoder acting as a universal interface for managing different modalities. Specifically, mPLUG-Owl2 incorporates shared functional modules to facilitate modality collaboration and introduces a modality-adaptive module that preserves modality-specific features. Extensive experiments reveal that mPLUG-Owl2 is capable of generalizing both text tasks and multi-modal tasks while achieving state-of-the-art performances with a single generalized model. Notably, mPLUG-Owl2 is the first MLLM model that demonstrates the modality collaboration phenomenon in both pure-text and multi-modal scenarios, setting a pioneering path in the development of future multi-modal foundation models.

Misinformation is a prevalent societal issue due to its potential high risks. Out-Of-Context (OOC) misinformation, where authentic images are repurposed with false text, is one of the easiest and most effective ways to mislead audiences. Current methods focus on assessing image-text consistency but lack convincing explanations for their judgments, which are essential for debunking misinformation. While Multimodal Large Language Models (MLLMs) have rich knowledge and innate capability for visual reasoning and explanation generation, they still lack sophistication in understanding and discovering the subtle cross-modal differences. In this paper, we introduce SNIFFER, a novel multimodal large language model specifically engineered for OOC misinformation detection and explanation. SNIFFER employs two-stage instruction tuning on InstructBLIP. The first stage refines the model's concept alignment of generic objects with news-domain entities and the second stage leverages OOC-specific instruction data generated by language-only GPT-4 to fine-tune the model's discriminatory powers. Enhanced by external tools and retrieval, SNIFFER not only detects inconsistencies between text and image but also utilizes external knowledge for contextual verification. Our experiments show that SNIFFER surpasses the original MLLM by over 40\% and outperforms state-of-the-art methods in detection accuracy. SNIFFER also provides accurate and persuasive explanations as validated by quantitative and human evaluations.

3D visual grounding plays a crucial role in scene understanding, with extensive applications in AR/VR. Despite the significant progress made in recent methods, the requirement of dense textual descriptions for each individual object, which is time-consuming and costly, hinders their scalability. To mitigate reliance on text annotations during training, researchers have explored language-free training paradigms in the 2D field via explicit text generation or implicit feature substitution. Nevertheless, unlike 2D images, the complexity of spatial relations in 3D, coupled with the absence of robust 3D visual language pre-trained models, makes it challenging to directly transfer previous strategies. To tackle the above issues, in this paper, we introduce a language-free training framework for 3D visual grounding. By utilizing the visual-language joint embedding in 2D large cross-modality model as a bridge, we can expediently produce the pseudo-language features by leveraging the features of 2D images which are equivalent to that of real textual descriptions. We further develop a relation injection scheme, with a Neighboring Relation-aware Modeling module and a Cross-modality Relation Consistency module, aiming to enhance and preserve the complex relationships between the 2D and 3D embedding space. Extensive experiments demonstrate that our proposed language-free 3D visual grounding approach can obtain promising performance across three widely used datasets - ScanRefer, Nr3D and Sr3D. Our codes are available at https://github.com/xibi777/3DLFVG

Recent trends in Large Vision Language Models (LVLMs) research have been increasingly focusing on advancing beyond general image understanding towards more nuanced, object-level referential comprehension. In this paper, we present and delve into the self-consistency capability of LVLMs, a crucial aspect that reflects the models' ability to both generate informative captions for specific objects and subsequently utilize these captions to accurately re-identify the objects in a closed-loop process. This capability significantly mirrors the precision and reliability of fine-grained visual-language understanding.Our findings reveal that the self-consistency level of existing LVLMs falls short of expectations, posing limitations on their practical applicability and potential. To address this gap, we introduce a novel fine-tuning paradigm named \textbf{Self-Consistency Tuning (SC-Tune)}. It features the synergistic learning of a cyclic describer-locator system. This paradigm is not only data-efficient but also exhibits generalizability across multiple LVLMs. Through extensive experiments, we demonstrate that SC-Tune significantly elevates performance across a spectrum of object-level vision-language benchmarks and maintains competitive or improved performance on image-level vision-language benchmarks. Both our model and code will be publicly available.

When we look around and perform complex tasks, how we see and selectively process what we see is crucial. However, the lack of this visual search mechanism in current multimodal LLMs (MLLMs) hinders their ability to focus on important visual details, especially when handling high-resolution and visually crowded images. To address this, we introduce V$^*$, an LLM-guided visual search mechanism that employs the world knowledge in LLMs for efficient visual querying. When combined with an MLLM, this mechanism enhances collaborative reasoning, contextual understanding, and precise visual grounding. This integration results in a new MLLM meta-architecture, named **S**how, S**EA**rch, and Tel**L** (SEAL). We further create V$^*$Bench, a benchmark specifically designed to evaluate MLLMs in their ability to process high-resolution images and focus on visual details. Our study highlights the necessity of incorporating visual search capabilities into multimodal systems. The code is available at https://github.com/penghao-wu/vstar

Vision-and-language models trained to match images with text can be combined with visual explanation methods to point to the locations of specific objects in an image. Our work shows that the localization --``grounding''-- abilities of these models can be further improved by finetuning for self-consistent visual explanations. We propose a strategy for augmenting existing text-image datasets with paraphrases using a large language model, and SelfEQ, a weakly-supervised strategy on visual explanation maps for paraphrases that encourages self-consistency. Specifically, for an input textual phrase, we attempt to generate a paraphrase and finetune the model so that the phrase and paraphrase map to the same region in the image. We posit that this both expands the vocabulary that the model is able to handle, and improves the quality of the object locations highlighted by gradient-based visual explanation methods (e.g. GradCAM). We demonstrate that SelfEQ improves performance on Flickr30k, ReferIt, and RefCOCO+ over a strong baseline method and several prior works. Particularly, comparing to other methods that do not use any type of box annotations, we obtain 84.07% on Flickr30k (an absolute improvement of 4.69%), 67.40% on ReferIt (an absolute improvement of 7.68%), and 75.10%, 55.49% on RefCOCO+ test sets A and B respectively (an absolute improvement of 3.74% on average).

The recent advance in vision-language models is largely attributed to the abundance of image-text data. We aim to replicate this success for video-language models, but there simply is not enough human-curated video-text data available. We thus resort to fine-tuning a video-language model from a strong image-language baseline with synthesized instructional data. The resulting video-language model is then used to auto-label millions of videos to generate high-quality captions. We show the adapted video-language model performs well on a wide range of video-language benchmarks. For instance, it surpasses the best prior result on open-ended NExT-QA by 2.8%. Besides, our model generates detailed descriptions for previously unseen videos, which provide better textual supervision than existing methods. Experiments show that a video-language dual-encoder model contrastively trained on these auto-generated captions is 3.8% better than the strongest baseline that also leverages vision-language models. Our best model outperforms state-of-the-art methods on MSR-VTT zero-shot text-to-video retrieval by 6%.

We present DenseAV, a novel dual encoder grounding architecture that learns high-resolution, semantically meaningful, and audio-visual aligned features solely through watching videos. We show that DenseAV can discover the "meaning" of words and the "location" of sounds without explicit localization supervision. Furthermore, it automatically discovers and distinguishes between these two types of associations without discriminative supervision. We show that our high-quality localization abilities arise from a new multi-head feature aggregation operator that directly compares dense image and audio representations for contrastive learning. In contrast, many other systems that learn "global" audio and video representations do not show high quality localization of words and sound. Finally, we contribute two new datasets to improve the evaluation of AV representations through speech and sound prompted semantic segmentation. On these and other datasets we show DenseAV dramatically outperforms the prior art on speech and sound prompted semantic segmentation. DenseAV outperforms the current state-of-the-art, ImageBind, on cross-modal retrieval using fewer than half of the parameters.

Significant advancements have been made in image editing with the recent advance of the Diffusion model. However, most of the current methods primarily focus on global or subject-level modifications, and often face limitations when it comes to editing specific objects when there are other objects coexisting in the scene, given solely textual prompts. In response to this challenge, we introduce an object-level generative task called Referring Image Editing (RIE), which enables the identification and editing of specific source objects in an image using text prompts. To tackle this task effectively, we propose a tailored framework called ReferDiffusion. It aims to disentangle input prompts into multiple embeddings and employs a mixed-supervised multi-stage training strategy. To facilitate further research in this domain, we introduce the RefCOCO-Edit dataset, comprising images, editing prompts, source object segmentation masks, and reference edited images for training and evaluation. Our extensive experiments demonstrate the effectiveness of our approach in identifying and editing target objects, while conventional general image editing and region-based image editing methods have difficulties in this challenging task.

In the pursuit of robust and generalizable environment perception and language understanding, the ubiquitous challenge of dataset bias continues to plague vision-and-language navigation (VLN) agents, hindering their performance in unseen environments. This paper introduces the generalized cross-modal causal transformer (GOAT), a pioneering solution rooted in the paradigm of causal inference. By delving into both observable and unobservable confounders within vision, language, and history, we propose the back-door and front-door adjustment causal learning (BACL and FACL) modules to promote unbiased learning by comprehensively mitigating potential spurious correlations. Additionally, to capture global confounder features, we propose a cross-modal feature pooling (CFP) module supervised by contrastive learning, which is also shown to be effective in improving cross-modal feature learning during pre-training. Extensive experiments across multiple VLN datasets (R2R, REVERIE, RxR, and SOON) underscore the superiority of our proposed method over previous state-of-the-art approaches. Code is available at https://github.com/CrystalSixone/VLN-GOAT.

Recent advancements in large video-language models have shown promising results in the field of video understanding.Recent advances in large video-language models have displayed promising outcomes in video comprehension. Current approaches straightforwardly convert video into language tokens and employ large language models for multi-modal tasks.However, this method often leads to the generation of irrelevant content, commonly known as ''hallucination'', as the length of the text increases and the impact of the video diminishes.To address this problem, we propose Vista-LLaMA , a novel framework that maintains the consistent distance between all visual tokens and any language tokens, irrespective of the generated text length.Vista-LLaMA omits relative position encoding when determining attention weights between visual and text tokens, retaining the position encoding for text and text tokens. This amplifies the effect of visual tokens on text generation, especially when the relative distance is longer between visual and text tokens. The proposed attention mechanism significantly reduces the chance of producing irrelevant text related to the video content.Furthermore, we present a sequential visual projector that projects the current video frame into tokens of language space with the assistance of the previous frame. This approach not only captures the temporal relationship within the video, but also allows less visual tokens to encompass the entire video.Our approach significantly outperforms various previous methods (e.g., Video-ChatGPT, MovieChat) on four challenging open-ended video question answering benchmarks. We reach an accuracy of 60.7 on the zero-shot NExT-QA and 60.5 on the zero-shot MSRVTT-QA, setting a new state-of-the-art performance.

This paper focuses on open-ended video question answering, which aims to find the correct answers from a large answer set in response to a video-related question. This is essentially a multi-label classification task, since a question may have multiple answers. However, due to annotation costs, the labels in existing benchmarks are always extremely insufficient, typically one answer per question. As a result, existing works tend to directly treat all the unlabeled answers as negative labels, leading to limited ability for generalization. In this work, we introduce a simple yet effective ranking distillation framework (RADI) to mitigate this problem without additional manual annotation. RADI employs a teacher model trained with incomplete labels to generate rankings for potential answers, which contain rich knowledge about label priority as well as label-associated visual cues, thereby enriching the insufficient labeling information. To avoid overconfidence in the imperfect teacher model, we further present two robust and parameter-free ranking distillation approaches: a pairwise approach which introduces adaptive soft margins to dynamically refine the optimization constraints on various pairwise rankings, and a listwise approach which adopts sampling-based partial listwise learning to resist the bias in teacher ranking. Extensive experiments on five popular benchmarks consistently show that both our pairwise and listwise RADIs outperform state-of-the-art methods. Further analysis demonstrates the effectiveness of our methods on the insufficient labeling problem.

Existing open-vocabulary image segmentation methods require a fine-tuning step on mask labels and/or image-text datasets. Mask labels are labor-intensive, which limits the number of categories in segmentation datasets. Consequently, the vocabulary capacity of pre-trained VLMs is severely reduced after fine-tuning. However, without fine-tuning, VLMs trained under weak image-text supervision tend to make suboptimal mask predictions.To alleviate these issues, we introduce a novel recurrent framework that progressively filters out irrelevant texts and enhances mask quality without training efforts. The recurrent unit is a two-stage segmenter built upon a frozen VLM. Thus, our model retains the VLM's broad vocabulary space and equips it with segmentation ability.Experiments show that our method outperforms not only the training-free counterparts, but also those fine-tuned with millions of data samples, and sets the new state-of-the-art records for both zero-shot semantic and referring segmentation. Concretely, we improve the current record by 28.8, 16.0, and 6.9 mIoU on Pascal VOC, COCO Object, and Pascal Context.

The gigapixel scale of whole slide images (WSIs) poses a challenge for histopathology multi-modal chatbots, requiring a global WSI analysis for diagnosis, compounding evidence from different WSI patches. Current visual instruction datasets, generated through large language models, focus on creating question/answer pairs for individual image patches, which may lack diagnostic capacity on their own in histopathology, further complicated by the absence of spatial grounding in histopathology image captions. To bridge this gap, we introduce Quilt-Instruct, a large-scale dataset of $107,131$ histopathology-specific instruction question/answer pairs, that is collected by leveraging educational histopathology videos from YouTube, which provides spatial localization of captions by automatically extracting narrators' cursor movements. In addition, we provide contextual reasoning by extracting diagnosis and supporting facts from the entire video content to guide the extrapolative reasoning of GPT-4. Using Quilt-Instruct, we train Quilt-LLaVA, which can reason beyond the given single image patch, enabling diagnostic reasoning and the capability of spatial awareness. To evaluate Quilt-LLaVA, we propose a comprehensive evaluation dataset created from $985$ images and $1283$ human-generated question-answers. We also thoroughly evaluate Quilt-LLaVA using public histopathology datasets, where Quilt-LLaVA significantly outperforms SOTA by over 10\% on relative GPT-4 score and 4\% and 9\% on open and closed set VQA. Our code, data, and model will be made publicly available.

Vision foundation models have been explored recently to build general-purpose vision systems. However, predominant paradigms, driven by casting instance-level tasks as an object-word alignment, bring heavy cross-modality interaction, which is not effective in prompting object detection and visual grounding. Another line of work that focuses on pixel-level tasks often encounters a large annotation gap of things and stuff, and suffers from mutual interference between foreground-object and background-class segmentation. In stark contrast to the prevailing methods, we present APE, a universal visual perception model for aligning and prompting everything all at once in an image to perform diverse tasks, i.e., detection, segmentation, and grounding, as an instance-level sentence-object matching paradigm. Specifically, APE advances the convergence of detection and grounding by reformulating language-guided grounding as open-vocabulary detection, which efficiently scales up model prompting to thousands of category vocabularies and region descriptions while maintaining the effectiveness of cross-modality fusion. To bridge the granularity gap of different pixel-level tasks, APE equalizes semantic and panoptic segmentation to proxy instance learning by considering any isolated regions as individual instances. APE aligns vision and language representation on broad data with natural and challenging characteristics all at once without task-specific fine-tuning. The extensive experiments on over 160 datasets demonstrate that, with only one-suit of weights, APE outperforms (or is on par with) the state-of-the-art models, proving that an effective yet universal perception for anything aligning and prompting is indeed feasible. Codes and trained models will be released.

We study visually grounded VideoQA in response to the emerging trends of utilizing pretraining techniques for video-language understanding. Specifically, by forcing vision-language models (VLMs) to answer questions and simultaneously provide visual evidence, we seek to ascertain the extent to which the predictions of such techniques are genuinely anchored in relevant video content, versus spurious correlations from language or irrelevant visual context. Towards this, we construct NExT-GQA -- an extension of NExT-QA with 10.5$K$ temporal grounding (or location) labels tied to the original QA pairs. With NExT-GQA, we scrutinize a series of state-of-the-art VLMs. Through post-hoc attention analysis, we find that these models are extremely weak in substantiating the answers despite their strong QA performance. This exposes the limitation of current VLMs in making reliable predictions. As a remedy, we further explore and propose a grounded-QA method via Gaussian mask optimization and cross-modal learning. Experiments with different backbones demonstrate that this grounding mechanism improves both grounding and QA. Our dataset and code will be released. With these efforts, we aim to push towards trustworthy VLMs in VQA systems.

This study targets a critical aspect of multi-modal LLMs' (LLMs\&VLMs) inference: explicit controllable text generation. Multi-modal LLMs empower multi-modality understanding with the capability of semantic generation yet bring less explainability and heavier reliance on prompt contents due to their autoregressive generative nature. While manipulating prompt formats could improve outputs, designing specific and precise prompts per task can be challenging and ineffective. To tackle this issue, we introduce a novel inference method, Prompt Highlighter, which enables users to highlight specific prompt spans to interactively control the focus during generation. Motivated by the classifier-free diffusion guidance, we form regular and unconditional context pairs based on highlighted tokens, demonstrating that the autoregressive generation in models can be guided in a classifier-free way. Notably, we find that, during inference, guiding the models with highlighted tokens through the attention weights leads to more desired outputs. Our approach is compatible with current LLMs and VLMs, achieving impressive customized generation results without training. Experiments confirm its effectiveness in focusing on input contexts and generating reliable content. Without tuning on LLaVA-v1.5, our method secured 70.7 in the MMBench test and 1552.5 in MME-perception. Code is available at https://github.com/dvlab-research/Prompt-Highlighter

Composed image retrieval (CIR) task takes a composed query of image and text, aiming to search relative images for both conditions. Conventional CIR approaches need a training dataset composed of triplets of query image, query text, and target image, which is very expensive to collect. Several recent works have worked on the zero-shot (ZS) CIR paradigm to tackle the issue without using pre-collected triplets. However, the existing ZS-CIR methods show limited backbone scalability and generalizability due to the lack of diversity of the input texts during training. We propose a novel CIR framework, only using language for its training. Our LoCIR (Language-only training for CIR) can be trained only with text datasets by a novel self-supervision named self-masking projection (SMP). We project the text latent embedding to the token embedding space and construct a new text by replacing the keyword tokens of the original text. Then, we let the new and original texts have the same latent embedding vector. With this simple strategy, LoCIR is surprisingly efficient and highly effective; LoCIR with CLIP ViT-G backbone is trained in 48 minutes and shows the best ZS-CIR performances on four different CIR benchmarks, CIRCO, GeneCIS, FashionIQ, and CIRR, even outperforming supervised method on FashionIQ.

This paper addresses the task of video question answering (videoQA) via a decomposed multi-stage, modular reasoning framework. Previous modular methods have shown promise with a single planning stage ungrounded in visual content. However, through a simple and effective baseline, we find that such systems can lead to brittle behavior in practice for challenging videoQA settings. Thus, unlike traditional single-stage planning methods, we propose a multi-stage system consisting of an event parser, a grounding stage, and a final reasoning stage in conjunction with an external memory. All stages are training-free, and performed using few-shot prompting of large models, creating interpretable intermediate outputs at each stage. By decomposing the underlying planning and task complexity, our method, MoReVQA, improves over prior work on standard videoQA benchmarks (NExT-QA, iVQA, EgoSchema, and ActivityNet-QA) with state-of-the-art results, and extensions to related tasks (grounded videoQA, paragraph captioning).

Chain-of-Thought (CoT) guides large language models (LLMs) to reason step-by-step, and can motivate their logical reasoning ability. While effective for logical tasks, CoT is not conducive to creative problem-solving which often requires out-of-box thoughts and is crucial for innovation advancements. In this paper, we explore the Leap-of-Thought (LoT) abilities within LLMs — a non-sequential, creative paradigm involving strong associations and knowledge leaps. To this end, we study LLMs on the popular Oogiri game which needs participants to have good creativity and strong associative thinking for responding unexpectedly and humorously to the given image, text, or both, and thus is suitable for LoT study. Then to investigate LLMs' LoT ability in the Oogiri game, we first build a multimodal and multilingual Oogiri-GO dataset which contains over 130,000 samples from the Oogiri game, and observe the insufficient LoT ability or failures of most existing LLMs on the Oogiri game. Accordingly, we introduce a creative Leap-of-Thought (CLoT) paradigm to improve LLM's LoT ability. CLoT first formulates the Oogiri-GO dataset into LoT-oriented instruction tuning data to train pretrained LLM for achieving certain LoT humor generation and discrimination abilities. Then CLoT designs an explorative self-refinement that encourages the LLM to generate more creative LoT data via exploring parallels between seemingly unrelated concepts and selects high-quality data to train itself for self-refinement. CLoT not only excels in humor generation in the Oogiri game as shown in Fig. 1 but also boosts creative abilities in various tasks like "cloud guessing game" and "divergent association task". These findings advance our understanding and offer a pathway to improve LLMs' creative capacities for innovative applications across domains. The dataset, code, and models have been released online: https://zhongshsh.github.io/CLoT.

Leveraging large language models (LLMs) to integrate off-the-shelf tools (e.g., visual models and image processing functions) is a promising research direction to build powerful visual assistants for solving diverse visual tasks. However, the learning capability is rarely explored in existing methods, as they freeze the used tools after deployment, thereby limiting the generalization to new environments requiring specific knowledge. In this paper, we propose CLOVA, a Closed-LOop Visual Assistant to address this limitation, which encompasses inference, reflection, and learning phases in a closed-loop framework. During inference, LLMs generate programs and execute corresponding tools to accomplish given tasks. The reflection phase introduces a multimodal global-local reflection scheme to analyze whether and which tool needs to be updated based on environmental feedback. Lastly, the learning phase uses three flexible manners to collect training data in real-time and introduces a novel prompt tuning scheme to update the tools, enabling CLOVA to efficiently learn specific knowledge for new environments without human involvement. Experiments show that CLOVA outperforms tool-usage methods by 5\% in visual question answering and multiple-image reasoning tasks, by 10\% in knowledge tagging tasks, and by 20\% in image editing tasks, highlighting the significance of the learning capability for general visual assistants.

3D visual grounding is a challenging task that often requires direct and dense supervision, notably the semantic label for each object in the scene. In this paper, we instead study the naturally supervised setting that learns from only 3D scene and QA pairs, where prior works underperform. We propose the Language-Regularized Concept Learner (LARC), which uses constraints from language as regularization to significantly improve the accuracy of neuro-symbolic concept learners in the naturally supervised setting. Our approach is based on two core insights: the first is that language constraints (e.g., a word's relation to another) can serve as effective regularization for structured representations in neuro-symbolic models; the second is that we can query large language models to distill such constraints from language properties. We show that LARC improves performance of prior works in naturally supervised 3D visual grounding, and demonstrates a wide range of 3D visual reasoning capabilities—from zero-shot composition, to data efficiency and transferability. Our method represents a promising step towards regularizing structured visual reasoning frameworks with language-based priors, for learning in settings without dense supervision.

Vision language models (VLM) have demonstrated remarkable performance across various downstream tasks. However, understanding fine-grained visual-linguistic concepts, such as attributes and inter-object relationships, remains a significant challenge. While several benchmarks aim to evaluate VLMs in finer granularity, their primary focus remains on the linguistic aspect, neglecting the visual dimension. Here, we highlight the importance of evaluating VLMs from both a textual and visual perspective. We introduce a progressive pipeline to synthesize images that vary in a specific attribute while ensuring consistency in all other aspects. Utilizing this data engine, we carefully design a benchmark, SPEC, to diagnose the comprehension of object size, position, existence, and count. Subsequently, we conduct a thorough evaluation of four leading VLMs on SPEC. Surprisingly, their performance is close to random guesses, revealing significant limitations. With this in mind, we propose a simple yet effective approach to optimize VLMs in fine-grained understanding, achieving significant improvements on SPEC without compromising the zero-shot performance. Results on two additional fine-grained benchmarks also show consistent improvements, further validating the transferability of our approach. Code and data are available at https://github.com/wjpoom/SPEC.

Graphical User Interface (GUI) automation holds significant promise for assisting users with complex tasks, thereby boosting human productivity. Existing works leveraging Large Language Model (LLM) or LLM-based AI agents have shown capabilities in automating tasks on Android and Web platforms. However, these tasks are primarily aimed at simple device usage and entertainment operations. This paper presents a novel benchmark, AssistGUI, to evaluate whether models are capable of manipulating the mouse and keyboard on the Windows platform in response to user-requested tasks. We carefully collected a set of 100 tasks from nine widely-used software applications, such as, After Effects and MS Word, each accompanied by the necessary project files for better evaluation. Moreover, we propose a multi-agent collaboration framework, which incorporates four agents to perform task decomposition, GUI parsing, action generation, and reflection. Our experimental results reveal that our multi-agent collaboration mechanism outshines existing methods in performance. Nevertheless, the potential remains substantial, with the best model attaining only a 46\% success rate on our benchmark. We conclude with a thorough analysis of the current methods' limitations, setting the stage for future breakthroughs in this domain.

Multimodal large language models (MLLMs), building upon the foundation of powerful large language models (LLMs), have recently demonstrated exceptional capabilities in generating not only texts but also images given interleaved multimodal inputs (acting like a combination of GPT-4V and DALL-E 3). However, existing MLLM benchmarks remain limited to assessing only models' comprehension ability of single image-text inputs, failing to keep up with the strides made in MLLMs. A comprehensive benchmark is imperative for investigating the progress and uncovering the limitations of current MLLMs. In this work, we categorize the capabilities of MLLMs into hierarchical levels from $L_0$ to $L_4$ based on the modalities they can accept and generate, and propose SEED-Bench, a comprehensive benchmark that evaluates the hierarchical capabilities of MLLMs. Specifically, SEED-Bench comprises 24K multiple-choice questions with accurate human annotations, which spans 27 dimensions, including the evaluation of both text and image generation. Multiple-choice questions with groundtruth options derived from human annotation enables an objective and efficient assessment of model performance, eliminating the need for human or GPT intervention during evaluation. We further evaluate the performance of 22 prominent open-source MLLMs and summarize valuable observations. By revealing the limitations of existing MLLMs through extensive evaluations, we aim for SEED-Bench to provide insights that will motivate future research towards the goal of General Artificial Intelligence.

We delve into Open Domain Generalization (ODG), marked by domain and category shifts between training's labeled source and testing's unlabeled target domains. Existing solutions to ODG face limitations due to constrained generalizations of traditional CNN backbones and errors in detecting target open samples in the absence of prior knowledge. Addressing these pitfalls, we introduce ODG-CLIP, harnessing the semantic prowess of the vision-language model, CLIP. Our framework brings forth three primary innovations:Firstly, distinct from prevailing paradigms, we conceptualize ODG as a multi-class classification challenge encompassing both known and novel categories. Central to our approach is modeling a unique prompt tailored for detecting unknown class samples, and to train this, we employ a readily accessible stable diffusion model, elegantly generating proxy images for the open class.Secondly, aiming for domain-tailored classification (prompt) weights while ensuring a balance of precision and simplicity, we devise a novel visual style-centric prompt learning mechanism.Finally, we infuse images with class-discriminative knowledge derived from the prompt space to augment the fidelity of CLIP's visual embeddings. We introduce a novel objective to safeguard the continuity of this infused semantic intel across domains, especially for the shared classes.Through rigorous testing on diverse datasets, covering closed and open-set DG contexts, ODG-CLIP demonstrates clear supremacy, consistently outpacing peers with performance boosts between 8\%-16\%.

The quality of the data and annotation upper-bounds the quality of a downstream model. While there exist large text corpora and image-text pairs, high-quality video-text data is much harder to collect. First of all, manual labeling is more time-consuming, as it requires an annotator to watch an entire video. Second, videos have a temporal dimension, consist of a number of scenes stacked together, and show multiple actions. Accordingly, to establish a video dataset with high-quality captions, we propose an automatic approach leveraging multimodal inputs, such as textual video description, subtitles, and individual video frames. Specifically, we curate 3.8M high-resolution videos from the publicly available HD-VILA-100M dataset. We then split them into semantically consistent video clips, and apply multiple cross-modality teacher models to obtain captions for each video. Next, we finetune a retrieval model on a small subset where the best caption of each video is manually selected and then employ the model in the whole dataset to select the best caption as the annotation. In this way, we get 70M videos paired with high-quality text captions. We dub the dataset as Panda-70M. We show the value of the proposed dataset on three downstream tasks: video captioning, video and text retrieval, and text-driven video generation. The models trained on the proposed data score substantially better on the majority of metrics across all the tasks.

Referring video segmentation relies on natural language expressions to identify and segment objects, often emphasizing motion clues. Previous works treat a sentence as a whole and directly perform identification at the video-level, mixing up static image-level cues with temporal motion cues. However, image-level features cannot well comprehend motion cues in sentences, and static cues are not crucial for temporal perception. In fact, static cues can sometimes interfere with temporal perception by overshadowing motion cues. In this work, we propose to decouple video-level referring expression understanding into static and motion perception, with a specific emphasis on enhancing temporal comprehension. Firstly, we introduce an expression-decoupling module to make static cues and motion cues perform their distinct role, alleviating the issue of sentence embeddings overlooking motion cues. Secondly, we propose a hierarchical motion perception module to capture temporal information effectively across varying timescales. Furthermore, we employ contrastive learning to distinguish the motions of visually similar objects. These contributions yield state-of-the-art performance across five datasets, including a remarkable 9.2% J&F improvement on the challenging MeViS dataset. Code is available at https://github.com/heshuting555/DsHmp.

While Multi-modal Language Models (\textit{MLMs}) demonstrate impressive multimodal ability, they still struggle on providing factual and precise responses for tasks like visual question answering (\textit{VQA}).In this paper, we address this challenge from the perspective of contextual information. We propose Causal Context Generation, \textbf{Causal-CoG}, which is a prompting strategy that engages contextual information to enhance precise VQA during inference. Specifically, we prompt MLMs to generate contexts, i.e, text description of an image, and engage the generated contexts for question answering. Moreover, we investigate the advantage of contexts on VQA from a causality perspective, introducing causality filtering to select samples for which contextual information is helpful. To show the effectiveness of Causal-CoG, we run extensive experiments on 10 multimodal benchmarks and show consistent improvements, \emph{e.g.}, +6.30\% on POPE, +13.69\% on Vizwiz and +6.43\% on VQAv2 compared to direct decoding, surpassing existing methods. We hope Casual-CoG inspires explorations of context knowledge in multimodal models, and serves as a plug-and-play strategy for MLM decoding.

We introduce Posterior Distillation Sampling (PDS), a novel optimization method for parametric image editing based on diffusion models. Existing optimization-based methods, which leverage the powerful 2D prior of diffusion models to handle various parametric images, have mainly focused on generation. Unlike generation, editing requires a balance between conforming to the target attribute and preserving the identity of the source content. Recent 2D image editing methods have achieved this balance by leveraging the stochastic latent encoded in the generative process of diffusion models. To extend the editing capabilities of diffusion models shown in pixel space to parameter space, we reformulate the 2D image editing method into an optimization form named PDS. PDS matches the stochastic latents of the source and the target, enabling the sampling of targets in diverse parameter spaces that align with a desired attribute while maintaining the source's identity. We demonstrate that this optimization resembles running a generative process with the target attribute, but aligning this process with the trajectory of the source's generative process. Extensive editing results in Neural Radiance Fields and Scalable Vector Graphics representations demonstrate that PDS is capable of sampling targets to fulfill the aforementioned balance across various parameter spaces.

The rapid advancement of large language models (LLMs) has accelerated the emergence of in-context learning (ICL) as a cutting-edge approach in the natural language processing domain. Recently, ICL has been employed in visual understanding tasks, such as semantic segmentation and image captioning, yielding promising results. However, existing visual ICL framework can not enable producing content across multiple modalities, which limits their potential usage scenarios. To address this issue, we present a new ICL framework for visual understanding with multi-modal output enabled. First, we quantize and embed both text and visual prompt into a unified representational space, structured as interleaved in-context sequences. Then a decoder-only sparse transformer architecture is employed to perform generative modeling on them, facilitating in-context learning. Thanks to this design, the model is capable of handling in-context vision understanding tasks with multimodal output in a unified pipeline. Experimental results demonstrate that our model achieves competitive performance compared with specialized models and previous ICL baselines. Overall, our research takes a further step toward unified multimodal in-context learning.

In the Visual Question Answering (VQA) task, recognizing and localizing entities pose significant challenges. Pretrained vision-and-language models have addressed this problem by providing a text description as the answer. However, in visual scenes with multiple entities, textual descriptions struggle to distinguish the entities from the same category effectively. Consequently, the VQA dataset is limited by the limitations of text description and cannot adequately cover scenarios involving multiple entities. To address this challenge, we introduce a module, Mask for Align (Mask4Align), which can determine the entity's position in the given image that best matches the user-input question. This module incorporates colored masks into the image, enabling the VQA model to handle discrimination and localization challenges associated with multiple entities. To process an arbitrary number of similar entities, Mask4Align is designed in a hierarchical manner to discern subtle differences, achieving precise localization. Furthermore, Mask4Align leverages pre-trained models and does not involve any extra task-sensitive parameters, and therefore, no additional training or hyper-parameter tuning is required. We evaluate our method on the gaze target prediction task and test our methods on different types of questions in the wild to verify the module's generalization. Code implementations are provided in the supplementary material.

Reasoning from visual dynamics scenes has many real-world applications. However, existing video reasoning benchmarks are still inadequate since they were mainly designed for factual or situated reasoning and rarely involve broader knowledge in the real world.Our work aims to delve deeper into reasoning evaluations, specifically within dynamic, open-world, and structured context knowledge. We propose a new benchmark (SOK-Bench), consisting of 44K questions and 10K situations with instance-level annotations depicted in the videos. The reasoning process is required to understand and apply situated knowledge and general knowledge for problem-solving.To create such a dataset, we propose an automatic and scalable generation method to generate question-answer pairs, knowledge graphs, and rationales by instructing the combinations of LLMs and MLLMs. Concretely, we first extract observable situated entities, relations, and processes from videos for situated knowledge and then extend to open-world knowledge beyond the visible content. The task generation is facilitated through multiple dialogues as iterations and subsequently corrected and refined by our designed self-promptings and demonstrations. With a corpus of both explicit situated facts and implicit commonsense, we generate associated question-answer pairs and reasoning processes, finally followed by manual reviews for quality assurance. We evaluated recent mainstream large vision-language models on the benchmark and found several insightful conclusions. For more information, please refer to our benchmark at www.bobbywu.com/SOKBench.

Despite the recent progress made in Video Question-Answering (VideoQA), these methods typically function as black-boxes, making it difficult to understand their reasoning processes and perform consistent compositional reasoning. To address these challenges, we propose a model-agnostic Video Alignment and Answer Aggregation (VA$^{3}$) framework, which is capable of enhancing both compositional consistency and accuracy of existing VidQA methods by integrating video aligner and answer aggregator modules. The video aligner hierarchically selects the relevant video clips based on the question, while the answer aggregator deduces the answer to the question based on its sub-questions, with compositional consistency ensured by the information flow along the question decompose graph and the contrastive learning strategy. We evaluate our framework on three settings of the AGQA-Decomp dataset with three baseline methods, and propose new metrics to measure the compositional consistency of VidQA methods more comprehensively. Moreover, we propose a large language model (LLM) based automatic question decompose pipeline to apply our framework on any VidQA data. We extend MSVD and NExT-QA datasets with it to evaluate such scheme and our VA$^3$ framework on broader scenarios. Extensive experiments show that our framework improves both compositional consistency and accuracy of existing methods, leading to more interpretable models in real-world applications.

We propose a method to efficiently equip the Segment Anything Model (SAM) with the ability to generate regional captions. SAM presents strong generalizability to segment anything while is short for semantic understanding. By introducing a lightweight query-based feature mixer, we align the region-specific features with the embedding space of language models for later caption generation. As the number of trainable parameters is small (typically in the order of tens of millions), it costs less computation, less memory usage, and less communication bandwidth, resulting in both fast and scalable training. To address the scarcity problem of regional caption data, we propose to first pre-train our model on objection detection and segmentation tasks. We call this step weak supervision pretraining since the pretraining data only contains category names instead of full-sentence descriptions. The weak supervision pretraining allows us to leverage many publicly available object detection and segmentation datasets. We conduct extensive experiments to demonstrate the superiority of our method and validate each design choice. This work serves as a stepping stone towards scaling up regional captioning data and sheds light on exploring efficient ways to augment SAM with regional semantics.

Hallucination, posed as a pervasive challenge of multi-modal large language models (MLLMs), has significantly impeded their real-world usage that demands precise judgment. Existing methods mitigate this issue with either training with specific designed data or inferencing with external knowledge from other sources, incurring inevitable additional costs. In this paper, we present $\textbf{OPERA}$, a novel MLLM decoding method grounded in an $\textbf{O}$ver-trust $\textbf{Pe}$nalty and a $\textbf{R}$etrospection-$\textbf{A}$llocation strategy, serving as a nearly $\textbf{free lunch}$ to alleviate the hallucination issue without additional data, knowledge, or training. Our approach begins with an interesting observation that, most hallucinations are closely tied to the knowledge aggregation patterns manifested in the self-attention matrix, i.e., MLLMs tend to generate new tokens by focusing on a few summary tokens, but not all the previous tokens. Such partial over-trust inclination results in the neglecting of image tokens and describes the image content with hallucination. Based on the observation, OPERA introduces a penalty term on the model logits during the beam-search decoding to mitigate the over-trust issue, along with a rollback strategy that retrospects the presence of summary tokens in the previously generated tokens, and re-allocate the token selection if necessary. With extensive experiments, OPERA shows significant hallucination-mitigating performance on different MLLMs and metrics, proving its effectiveness and generality. Our code is at: https://github.com/shikiw/OPERA.

Generative vision-language models (VLMs) have shown impressive performance on zero-shot vision-language tasks through generalization-based text generation. However, recent studies have improved upon this generalization capability for zero-shot VL tasks by utilizing second-stage instruction tuning with a large amount of human-labeled and externally generated data from large language models.In this work, we propose the image-conditioned text correction task for enhancing zero-shot text generation with unlabeled data. By leveraging the inherent structure of language, we produce the image-text pair containing mismatched concepts, and VLMs are required to identify and correct the error according to the vision modality by text generation.Our experiments demonstrate that our second-stage tuning framework significantly enhances the generalization capabilities of VLMs on various image-to-text generation-based VL tasks.This work represents a promising direction for advancing the field of zero-shot inference in VLMs by providing a cost-effective and scalable solution for enhancing generalization performance.

Traditional referring expression comprehension (REC) aims to locate the target referent in an image guided by a text query. Several previous methods have studied on the Counterfactual problem in REC (C-REC) where the objects for a given query cannot be found in the image. However, these methods focus on the overall image-text or specific attribute mismatch only. In this paper, we address the C-REC problem from a deep perspective of fine-grained attributes. To this aim, we first propose a fine-grained counterfactual sample generation method to construct C-REC datasets. Specifically, we leverage pre-trained language model such as BERT to modify the attribute words in the queries, obtaining the corresponding counterfactual samples. Furthermore, we propose a C-REC framework. We first adopt three encoders to extract image, text and attribute features. Then, our dual-branch attentive fusion module fuses these cross-modal features with two branches by an attention mechanism. At last, two prediction heads generate a bounding box and a counterfactual label, respectively. In addition, we incorporate contrastive learning with the generated counterfactual samples as negatives to enhance the counterfactual perception. Extensive experiments show that our framework achieves promising performance on both public REC datasets RefCOCO/+/g and our constructed C-REC datasets C-RefCOCO/+/g. The code and data are available at https://github.com/Glacier0012/CREC.

Referring Expression Comprehension (REC) aims to localize the target objects specified by free-form natural language descriptions in images. While state-of-the-art methods achieve impressive performance, they perform dense perception of images, which incorporates redundant visual regions unrelated to linguistic queries, leading to additional computational overhead. This inspires us to explore a question: can we eliminate linguistic-irrelevant redundant visual regions to improve the efficiency of the model ? Existing relevant methods primarily focus on fundamental visual tasks, with limited exploration in vision-language fields. To address this, we propose a coarse-to-fine iterative perception framework, called ScanFormer. It can iteratively exploit the image scale pyramid to extract linguistic-relevant visual patches from top to bottom. In each iteration, irrelevant patches are discarded by our designed informativeness prediction. Furthermore, we propose a patch selection strategy for discarded patches to accelerate inference. Experiments on widely used datasets, namely RefCOCO, RefCOCO+, RefCOCOg, and ReferItGame, verify the effectiveness of our method, which can strike a balance between accuracy and efficiency.

Current open-source Large Multimodal Models (LMMs) excel at tasks such as open-vocabulary language grounding and segmentation but can suffer under false premises, when queries imply the existence of something that is not actually present in the image. We observe that existing methods which fine-tune a LMM to segment images significantly degrade their ability to reliably determine ("see") if an object is present, a form of catastrophic forgetting. In this work, we propose cascading and/or joint training LMMs to solve this task, avoiding catastrophic forgetting of previous skills. Our resulting model can "see" by detecting whether objects are present in an image, "say" by telling the user if they are not, proposing alternative queries or correcting semantic errors in the query, and finally``segment'' by outputting the mask of the desired objects if they exist. We introduce a novel dataset and benchmark of false premise queries for existing RefCOCO(+/g) data (which we call FP-RefCOCO(+/g)), and demonstrate that our method not only detects false premises up to 55\% better than existing approaches, but under false premise conditions produces relative cIOU improvements of more than 31\% over baselines, and produces natural language feedback judged helpful up to 67\% of the time.

Despite the recent success of sign language research, the widely adopted CNN-based backbones are mainly migrated from other computer vision tasks, in which the contours and texture of objects are crucial for {identifying} objects. They usually treat sign frames as grids and may fail to capture effective cross-region features. In fact, sign language tasks need to focus on the correlation of different regions in one frame and the interaction of different regions among adjacent frames for {identifying} a sign sequence. In this paper, we propose to represent a sign sequence as graphs and introduce a simple yet effective graph-based sign language processing architecture named SignGraph, to extract cross-region features at the graph level. SignGraph consists of two basic modules: Local Sign Graph ($LSG$) module for learning the correlation of \textbf{intra-frame cross-region} features in one frame and Temporal Sign Graph ($TSG$) module for tracking the interaction of \textbf{inter-frame cross-region} features among adjacent frames. With $LSG$ and $TSG$, we build our model in a multiscale manner to ensure that the representation of nodes can capture cross-region features at different granularities. Extensive experiments on current public sign language datasets demonstrate the superiority of our SignGraph model. Our model achieves very competitive performances with the SOTA model, while not using any extra cues. Code and models are available at: \href{https://github.com/gswycf/SignGraph}{\textcolor{blue}{https://github.com/gswycf/SignGraph}}.

We propose Strongly Supervised pre-training with ScreenShots (S4) - a novel pre-training paradigm for Vision-Language Models using data from large-scale web screenshot rendering. Using web screenshot unlock a treasure trove of visual and textual cues that are simply not present in using image-text pairs. In S4, we leverage the inherent tree-structure hierarchy of HTML elements and the spatial localization to carefully design 10 pre-training tasks with large scale annotated data. These tasks resembles downstream tasks across different domains and the annotations are cheap to obtain. We demonstrate that, comparing to current screenshot pre-training objectives, our innovative pre-training method significantly enhances performance of image-to-text model in nine varied and popular downstream tasks - up to 76.1% improvements on Table Detection, and at least 1% on Widget Captioning.

We demonstrate text as a strong cross-modal interface. Rather than relying on deep embeddings to connect image and language as the interface representation, our approach represents an image as text, from which we enjoy the interpretability and flexibility inherent to natural language. We employ an autoencoder that uses a pre-trained text-to-image diffusion model for decoding. The encoder is trained to transform an input image into text, which is then fed into the fixed text-to-image diffusion decoder to reconstruct the original input -- a process we term De-Diffusion. Experiments validate both the precision and comprehensiveness of De-Diffusion text representing images, such that it can be readily ingested by off-the-shelf text-to-image tools and LLMs for diverse multi-modal tasks. For example, a single De-Diffusion model can generalize to provide transferable prompts for different text-to-image tools, and also achieves a new state of the art on open-ended vision-language tasks by simply prompting LLMs with few-shot examples. Project page: https://dediffusion.github.io/

With the success of large language models (LLMs), integrating the vision model into LLMs to build vision-language foundation models has gained much more interest recently. However, existing LLM-based large multimodal models (e.g., Video-LLaMA, VideoChat) can only take in a limited number of frames for short video understanding. In this study, we mainly focus on designing an efficient and effective model for long-term video understanding. Instead of trying to process more frames simultaneously like most existing work, we propose to process videos in an online manner and store past video information in a memory bank. This allows our model to reference historical video content for long-term analysis without exceeding LLMs' context length constraints or GPU memory limits. Our memory bank can be seamlessly integrated into current multimodal LLMs in an off-the-shelf manner. We conduct extensive experiments on various video understanding tasks, such as long-video understanding, video question answering, and video captioning, and our model can achieve state-of-the-art performances across multiple datasets.

Existing object recognition models have been shown to lack robustness in diverse geographical scenarios due to domain shifts in design and context. Class representations need to be adapted to more accurately reflect an object concept under these shifts. In the absence of training data from target geographies, we hypothesize that geographically diverse descriptive knowledge of categories can enhance robustness. For this purpose, we explore the feasibility of probing a large language model for geography-based object knowledge, and we examine the effects of integrating knowledge into zero-shot and learnable soft prompting with CLIP. Within this exploration, we propose geography knowledge regularization to ensure that soft prompts trained on a source set of geographies generalize to an unseen target set. Accuracy gains over prompting baselines on DollarStreet while training only on Europe data are up to +2.8/1.2/1.6 on target data from Africa/Asia/Americas, and +4.6 overall on the hardest classes. Competitive performance is shown vs. few-shot target training, and analysis is provided to direct future study of geographical robustness.

Understanding human actions from videos of first-person view poses significant challenges. Most prior approaches explore representation learning on egocentric videos only, while overlooking the potential benefit of exploiting existing large-scale third-person videos. In this paper, (1) we develop EgoInstructor, a retrieval-augmented multimodal captioning model that automatically retrieves semantically relevant third-person instructional videos to enhance the video captioning of egocentric videos. (2) For training the cross-view retrieval module, we devise an automatic pipeline to discover ego-exo video pairs from distinct large-scale egocentric and exocentric datasets. (3) We train the cross-view retrieval module with a novel EgoExoNCE loss that pulls egocentric and exocentric video features closer by aligning them to shared text features that describe similar actions. (4) Through extensive experiments, our cross-view retrieval module demonstrates superior performance across seven benchmarks. Regarding egocentric video captioning, EgoInstructor exhibits significant improvements by leveraging third-person videos as references.

Vision-language (VL) models have achieved unprecedented success recently, in which the connection module is the key to bridge the modality gap. Nevertheless, the abundant visual clues are not sufficiently exploited in most existing methods. On the vision side, most existing approaches only use the last feature of the vision tower, without using the low-level features. On the language side, most existing methods only introduce shallow vision-language interactions. In this paper, we present a vision-inspired vision-language connection module, dubbed as VIVL, which efficiently exploits the vision cue for VL models. To take advantage of the lower-level information from the vision tower, a feature pyramid extractor (FPE) is introduced to combine features from different intermediate layers, which enriches the visual cue with negligible parameters and computation overhead. To enhance VL interactions, we propose deep vision-language prompts (DVLP) that allows deep interactions of vision and language features efficiently. Our VIVL exceeds the previous state-of-the-art method by 18.1 CIDEr when training from scratch on the COCO caption task, which greatly improves the data efficiency. When used as a plug-in module, VIVL consistently improves the performance for various backbones and VL frameworks, delivering new state-of-the-art results on multiple benchmarks, e.g., NoCaps and VQAv2.

Vision-Language Models (VLMs), such as Flamingo and GPT-4V, have shown immense potential by integrating large language models with vision systems. Nevertheless, these models face challenges in the fundamental computer vision task of object localisation, due to their training on multimodal data containing mostly captions without explicit spatial grounding. While it is possible to construct custom, supervised training pipelines with bounding box annotations that integrate with VLMs, these result in specialized and hard-to-scale models. In this paper, we aim to explore the limits of caption-based VLMs and instead propose to tackle the challenge in a simpler manner by i) keeping the weights of a caption-based VLM frozen and ii) not using any supervised detection data. To this end, we introduce an input-agnostic Positional Insert (PIN), a learnable spatial prompt, containing a minimal set of parameters that are slid inside the frozen VLM, unlocking object localisation capabilities. Our PIN module is trained with a simple next-token prediction task on synthetic data without requiring the introduction of new output heads. Our experiments demonstrate strong zero-shot localisation performances on a variety of images, including Pascal VOC, COCO, LVIS, and diverse images like paintings or cartoons.

Establishing an automatic evaluation metric that closely aligns with human judgments is essential for effectively developing image captioning models. Recent data-driven metrics have demonstrated a stronger correlation with human judgments than classic metrics such as CIDEr; however they lack sufficient capabilities to handle hallucinations and generalize across diverse images and texts partially because they compute scalar similarities merely using embeddings learned from tasks unrelated to image captioning evaluation.In this study, we propose Polos, a supervised automatic evaluation metric for image captioning models.Polos computes scores from multimodal inputs, using a parallel feature extraction mechanism that leverages embeddings trained through large-scale contrastive learning.To train Polos, we introduce Multimodal Metric Learning from Human Feedback (M$^2$LHF), a framework for developing metrics based on human feedback. We constructed the Polaris dataset, which comprises 131K human judgments from 550 evaluators, which is approximately ten times larger than standard datasets. Our approach achieved state-of-the-art performance on Composite, Flickr8K-Expert, Flickr8K-CF, PASCAL-50S, FOIL, and the Polaris dataset, thereby demonstrating its effectiveness and robustness.

Video Paragraph Grounding (VPG) is an emerging task in video-language understanding, which aims at localizing multiple sentences with semantic relations and temporal order from an untrimmed video. However, existing VPG approaches are heavily reliant on a considerable number of temporal labels that are laborious and time-consuming to acquire. In this work, we introduce and explore Weakly-Supervised Video Paragraph Grounding (WSVPG) to eliminate the need of temporal annotations. Different from previous weakly-supervised grounding frameworks based on multiple instance learning or reconstruction learning for two-stage candidate ranking, we propose a novel siamese learning framework that jointly learns the cross-modal feature alignment and temporal coordinate regression without timestamp labels to achieve concise one-stage localization for WSVPG. Specifically, we devise a Siamese Grounding TRansformer (SiamGTR) consisting of two weight-sharing branches for learning complementary supervision. An Augmentation Branch is utilized for directly regressing the temporal boundaries of a complete paragraph within a pseudo video, and an Inference Branch is designed to capture the order-guided feature correspondence for localizing multiple sentences in a normal video. We demonstrate by extensive experiments that our paradigm has superior practicability and flexibility to achieve efficient weakly-supervised or semi-supervised learning, outperforming state-of-the-art methods trained with the same or stronger supervision.

Long video question answering is a challenging task that involves recognizing short-term activities and reasoning about their fine-grained relationships. State-of-the-art video Large Language Models (vLLMs) hold promise as a viable solution due to their demonstrated emergent capabilities on new tasks. However, despite being trained on millions of short seconds-long videos, vLLMs are unable to understand minutes-long videos and accurately answer questions about them. To address this limitation, we propose a lightweight and self-supervised approach, Key frame-conditioned long video-LLM (Koala), that introduces learnable spatiotemporal queries to adapt pretrained vLLMs for generalizing to longer videos. Our approach introduces two new tokenizers that condition on visual tokens computed from sparse video key frames for understanding short and long video moments. We train our proposed approach on HowTo100M and demonstrate its effectiveness on zero-shot long video understanding benchmarks, where it outperforms state-of-the-art large models by 3 - 6% in absolute accuracy across all tasks. Surprisingly, we also empirically show that our approach not only helps a pretrained vLLM to understand long videos but also improves its accuracy on short-term action recognition.

The recent progress in language-based open-vocabulary object detection can be largely attributed to finding better ways of leveraging large-scale data with free-form text annotations. Training such models with a discriminative objective function has proven successful, but requires good positive and negative samples. However, the free-form nature and the open vocabulary of object descriptions make the space of negatives extremely large. Prior works randomly sample negatives or use rule-based techniques to build them. In contrast, we propose to leverage the vast knowledge built into modern generative models to automatically build negatives that are more relevant to the original data. Specifically, we use large-language-models to generate negative text descriptions, and text-to-image diffusion models to also generate corresponding negative images. Our experimental analysis confirms the relevance of the generated negative data, and its use in language-based detectors improves performance on two complex benchmarks. Code is available at https://github.com/xiaofeng94/Gen-Enhanced-Negs.

Sequence-to-sequence vision-language models are showing promise, but their applicability is limited by their inference latency due to their autoregressive way of generating predictions.We propose a sequence-to-sequence vision-language model with a flexible hypothesis space, manifest in the training set and encoded in a layer of learnable query tokens. The architecture is trained with a novel loss, inspired by the language domain, that marginalizes over multiple inference paths in the decoder. This enables us the flexibility to adapt the hypothesis space to the task, rather than restricting to the embedding of a single token as in an autoregressive model. The resulting model, NARVL, achieves performance on-par with its autoregressive counterpart, but is faster at inference time since the decoder has to be executed once to jointly produce all output tokens, rather than sequentially to produce them one at a time. We test our model on four vision-language tasks, and perform ablation studies to single out the contribution of each component.

We introduce MMMU: a new benchmark designed to evaluate multimodal models on massive multi-discipline tasks demanding a fusion of visual and textual understanding with college-level reasoning. MMMU includes 11.5K meticulously collected multimodal questions from college exams, quizzes, and textbooks, covering six core disciplines: Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering. These questions span 30 subjects and 183 subfields, comprising 32 highly heterogeneous image types, such as charts, diagrams, maps, tables, music sheets, and chemical structures. Unlike existing benchmarks, MMMU focuses on advanced perception and reasoning with domain-specific knowledge, challenging models to perform tasks akin to those faced by experts. Our evaluation of 14 open-source LMMs and the proprietary GPT-4V(ision) highlights the substantial challenges posed by MMMU. Even the advanced GPT-4V only achieves a 56% accuracy, indicating significant room for improvement. We believe MMMU will stimulate the community to build next-generation multimodal foundation models towards expert artificial general intelligence.

Understanding data visualizations like charts and plots requires reasoning about both visual elements and numerics. Although strong in extractive questions, current chart visual question answering (chart VQA) models suffer on complex reasoning questions. In this work, we address the lack of reasoning ability by data augmentation. We leverage Large Language Models (LLMs), which have shown to have strong reasoning ability, as an automatic data annotator that generates question-answer annotations for chart images. The key innovation in our method lies in the Synthesize Step-by-Step strategy: our LLM-based data generator learns to decompose the complex question into step-by-step sub-questions (rationales), which are then used to derive the final answer using external tools, ie Python. This step-wise generation procedure is trained on synthetic data generated using a template-based QA generation pipeline. Experimental results highlight the significance of the proposed step-by-step generation. By training with the LLM-augmented data (LAMENDA), we significantly enhance the chart VQA models, achieving the state-of-the-art accuracy on the ChartQA and PlotQA datasets. In particular, our approach improves the accuracy of the previous state-of-the-art approach from 38% to 54% on the human-written questions in the ChartQA dataset, which needs strong reasoning. We hope our work underscores the potential of synthetic data and encourages further exploration of data augmentation using LLMs for reasoning-heavy tasks.

Building a generalist agent that can interact with the world is an ultimate goal for humans, thus spurring the research for embodied navigation, where an agent is required to navigate according to instructions or respond to queries. Despite the major progress attained, previous works primarily focus on task-specific agents and lack generalizability to unseen scenarios. Recently, LLMs have presented remarkable capabilities across various fields, and provided a promising opportunity for embodied navigation. Drawing on this, we propose the first generalist model for embodied navigation, NaviLLM. It adapts LLMs to embodied navigation by introducing schema-based instruction. The schema-based instruction flexibly casts various tasks into generation problems, thereby unifying a wide range of tasks. This approach allows us to integrate diverse data sources from various datasets into the training, equipping NaviLLM with a wide range of capabilities required by embodied navigation. We conduct extensive experiments to evaluate the performance and generalizability of our model. The experimental results demonstrate that our unified model achieves state-of-the-art performance on CVDN, SOON, and ScanQA. Specifically, it surpasses the previous stats-of-the-art method by a significant margin of 29\% in goal progress on CVDN. Moreover, our model also demonstrates strong generalizability and presents impressive results on unseen tasks, e.g., embodied question answering and 3D captioning.

We introduce multimodal story summarization by leveraging TV episode recaps – short video sequences interweaving key story moments from previous episodes to bring viewers up to speed. We propose PlotSnap, a dataset featuring two crime thriller TV shows with rich recaps and long episodes of 40 minutes. Story summarization labels are unlocked by matching recap shots to corresponding sub-stories in the episode. We propose a hierarchical model TaleSumm that processes entire episodes by creating compact shot and dialog representations, and predicts importance scores for each video shot and dialog utterance by enabling interactions between local story groups. Unlike traditional summarization, our method extracts multiple plot points from long videos. We present a thorough evaluation on story summarization, including promising cross-series generalization. TaleSumm also shows good results on classic video summarization benchmarks.

We present MM-Narrator, a novel system leveraging GPT-4 with multimodal in-context learning for the generation of audio descriptions (AD). Unlike previous methods that primarily focused on downstream fine-tuning with short video clips, MM-Narrator excels in generating precise audio descriptions for videos of extensive lengths, even beyond hours, in an autoregressive manner. This capability is made possible by the proposed memory-augmented generation process, which effectively utilizes both the short-term textual context and long-term visual memory through an efficient register-and-recall mechanism. These contextual memories compile pertinent past information, including storylines and character identities, ensuring an accurate tracking and depicting of story-coherent and character-centric audio descriptions. Maintaining the training-free design of MM-Narrator, we further propose a complexity-based demonstration selection strategy to largely enhance its multi-step reasoning capability via few-shot multimodal in-context learning (MM-ICL). Experimental results on MAD-eval dataset demonstrate that MM-Narrator consistently outperforms both the existing fine-tuning-based approaches and LLM-based approaches in most scenarios, as measured by standard evaluation metrics. Additionally, we introduce first segment-based evaluator for recurrent text generation. Empowered by GPT-4, this evaluator comprehensively reasons and marks the AD generation performance across various extendable dimensions.

The recent progress in Large Language Models (LLM) has spurred various advancements in image-language conversation agents, while how to build a proficient video-based dialogue system is still under exploration. Considering the extensive scale of LLM and visual backbone, minimal GPU memory is left for facilitating effective temporal modeling, which is crucial for comprehending and providing feedback on videos. To this end, we propose Branching Temporal Adapter (BT-Adapter), a novel method for extending image-language pretrained models into the video domain. Specifically, BT-Adapter serves as a plug-and-use temporal modeling branch alongside the pretrained visual encoder, which is tuned while keeping the backbone frozen. Just pretrained once, BT-Adapter can be seamlessly integrated into all image conversation models using this version of CLIP, enabling video conversations without the need for video instructions.%The BT-Adapter pretrained on video datasets can be seamlessly integrated with the image chatbots, enhancing the model's capabilities in temporal modeling and enabling video conversations without the need for video instructions.Besides, we develop a unique asymmetric token masking strategy inside the branch with tailor-made training tasks for BT-Adapter, facilitating faster convergence and better results.Thanks to BT-Adapter, we are able to empower existing multimodal dialogue models with strong video understanding capabilities without incurring excessive GPU costs.Without bells and whistles, BT-Adapter achieves (1) state-of-the-art zero-shot results on various video tasks using thousands of fewer GPU hours. (2) better performance than current video chatbots without any video instruction tuning. (3) state-of-the-art results of video chatting using video instruction tuning, outperforming previous SOTAs by a large margin.

The rise of multimodal large language models (MLLMs) has spurred interest in language-based driving tasks. However, existing research typically focuses on limited tasks and often omits key multi-view and temporal information which is crucial for robust autonomous driving. To bridge these gaps, we introduce NuInstruct, a novel dataset with 91K multi-view video-QA pairs across 17 subtasks, where each task demands holistic information (e.g., temporal, multi-view, and spatial), significantly elevating the challenge level. To obtain NuInstruct, we propose a novel SQL-based method to generate instruction-response pairs automatically, which is inspired by the driving logical progression of humans. We further present BEV-InMLLM, an end-to-end method for efficiently deriving instruction-aware Bird’s-Eye-View (BEV) features, language-aligned for large language models. BEV-InMLLM integrates multi-view, spatial awareness, and temporal semantics to enhance MLLMs' capabilities on NuInstruct tasks. Moreover, our proposed BEV injection module is a plug-and-play method for existing MLLMs. Our experiments on NuInstruct demonstrate that BEV-InMLLM significantly outperforms existing MLLMs, e.g. ~9% improvement on various tasks. We plan to release our \dataset~for future research development.

Being able to carry out complicated vision-language reasoning tasks in 3D space represents a significant milestone in developing household robots and human-centered embodied AI. In this work, we demonstrate that a critical and distinct challenge in 3D vision language reasoning is the situational awareness, which incorporates two key components: (1) The autonomous agent grounds its self-location based on a language prompt. (2) The agent answers open-ended questions from the perspective of its calculated position. To address this challenge, we propose SIG3D, an end-to-end Situation-Grounded model for 3D vision language reasoning. We tokenize the 3D scene into sparse voxel representation, and propose a language-grounded situation estimator followed by a situated question answering module. Experiments on the SQA3D and ScanQA datasets show that SIG3D outperforms state-of-the-art models in situational estimation and question answering by a larger margin (e.g., an enhancement of over 30% on situation accuracy). Subsequent analysis corroborates our architectural design choices, explores the distinct functions of visual and textual tokens, and highlights the importance of situational awareness in the domain of 3D question-answering.

Is vision good enough for language? Recent advancements in multimodal models primarily stem from the powerful reasoning abilities of large language models (LLMs). However, the visual component typically depends only on the instance-level contrastive language-image pre-training (CLIP). Our research reveals that the visual capabilities in recent Multimodal LLMs (MLLMs) still exhibit systematic shortcomings. To understand the roots of these errors, we explore the gap between the visual embedding space of CLIP and vision-only self-supervised learning. We identify ``CLIP-blind pairs'' – images that CLIP perceives as similar despite their clear visual differences. With these pairs, we construct the Multimodal Visual Patterns (MMVP) benchmark. MMVP exposes areas where state-of-the-art systems, including GPT-4V, struggle with straightforward questions across nine basic visual patterns, often providing incorrect answers and hallucinated explanations. We further evaluate various CLIP-based vision-and-language models and found a notable correlation between visual patterns that challenge CLIP models and those problematic for multimodal LLMs. As an initial effort to address these issues, we propose a Mixture of Features (MoF) approach, demonstrating that integrating vision self-supervised learning features with MLLMs can significantly enhance their visual grounding capabilities. Together, our research suggests visual representation learning remains an open challenge, and accurate visual grounding is crucial for future successful multimodal systems.

In recent years, large-scale video-language pre-training (VidLP) has received considerable attention for its effectiveness in relevant tasks. In this paper, we propose a novel action-centric VidLP framework that employs video tube features for temporal modeling and language features based on semantic role labeling (SRL). Our video encoder generates multiple tube features along object trajectories, identifying action-related regions within videos, to overcome the limitations of existing query-driven attention mechanisms. Additionally, our text encoder incorporates high-level, action-related language knowledge, previously underutilized in current VidLP models. The SRL captures action-verbs and related semantics among objects in sentences and enhances the ability to perform instance-level text matching, thus enriching the cross-modal (CM) alignment process. We also introduce two novel pre-training objectives and a self-supervision strategy to produce a more faithful CM representation. Experimental results demonstrate that our method outperforms existing VidLP frameworks in various downstream tasks and datasets, establishing our model a baseline in the modern VidLP framework.

Large language models have demonstrated impressive universal capabilities across a wide range of open-ended tasks and have extended their utility to encompass multimodal conversations. However, existing methods encounter challenges in effectively handling both image and video understanding, particularly with limited visual tokens. In this work, we introduce Chat-UniVi, a Unified Vision-language model capable of comprehending and engaging in conversations involving images and videos through a unified visual representation. Specifically, we employ a set of dynamic visual tokens to uniformly represent images and videos. This representation framework empowers the model to efficiently utilize a limited number of visual tokens to simultaneously capture the spatial details necessary for images and the comprehensive temporal relationship required for videos. Moreover, we leverage a multi-scale representation, enabling the model to perceive both high-level semantic concepts and low-level visual details. Notably, Chat-UniVi is trained on a mixed dataset containing both images and videos, allowing direct application to tasks involving both mediums without requiring any modifications. Extensive experimental results demonstrate that Chat-UniVi consistently outperforms even existing methods exclusively designed for either images or videos. Code is available at https://github.com/PKU-YuanGroup/Chat-UniVi.

Referring image segmentation (RIS) aims to precisely segment referents in images through corresponding natural language expressions, yet relying on cost-intensive mask annotations. Weakly supervised RIS thus learns from image-text pairs to pixel-level semantics, which is challenging for segmenting fine-grained masks. A natural approach to enhancing segmentation precision is to empower weakly supervised RIS with the image segmentation foundation model SAM. Nevertheless, we observe that simply integrating SAM yields limited benefits and can even lead to performance regression due to the inevitable noise issues and challenges in excessive focus on object parts. In this paper, we present an innovative framework, Point PrompTing (PPT), incorporated with the proposed multi-source curriculum learning strategy to address these challenges. Specifically, the core of PPT is a point generator that not only harnesses CLIP's text-image alignment capability and SAM's powerful mask generation ability but also generates negative point prompts to address the noisy and excessive focus issues inherently and effectively. In addition, we introduce a curriculum learning strategy with object-centric images to help PPT gradually learn from simpler yet precise semantic alignment to more complex RIS. Experiments demonstrate that our PPT significantly and consistently outperforms prior weakly supervised techniques on mIoU by 11.34\%, 14.14\%, and 6.97\% across RefCOCO, RefCOCO+, and G-Ref, respectively.

Visual prompting of large vision language models such as CLIP exhibit intriguing zero-shot capabilities. A manually drawn red circle, commonly used for highlighting, can guide CLIP's attention to the surrounding region, to identify specific objects within an image. Without precise object proposals, however, it is insufficient for localization. Our novel, simple yet effective approach enables CLIP to zero-shot localize: given an image and a text prompt describing an object, we first pick an initial ellipse from uniformly distributed anchor ellipses on the image grid via visual prompting, then use three loss functions to tune the ellipse coefficients to encapsulate the target region gradually. This yields promising experimental results for referring expression comprehension without precisely specified object proposals. In addition, we systematically present the limitations of visual prompting inherent in CLIP and discuss potential avenues for improvement. Code will be released.

Large language models (LLMs)-based image captioning has the capability of describing objects not explicitly observed in training data; yet novel objects occur frequently, necessitating the requirement of sustaining up-to-date object knowledge for open-world comprehension. Instead of relying on large amounts of data and/or scaling up network parameters, we introduce a highly effective retrieval-augmented image captioning method that prompts LLMs with object names retrieved from External Visual--name memory (EVCap). We build ever-changing object knowledge memory using objects' visuals and names, enabling us to (i) update the memory at a minimal cost and (ii) effortlessly augment LLMs with retrieved object names by utilizing a lightweight and fast-to-train model. Our model, which was trained only on the COCO dataset, can adapt to out-of-domain without requiring additional fine-tuning or re-training. Our experiments conducted on benchmarks and synthetic commonsense-violating data show that EVCap, with only 3.97M trainable parameters, exhibits superior performance compared to other methods based on frozen pre-trained LLMs. Its performance is also competitive to specialist SOTAs that require extensive training.

Diffusion models have shown tremendous results in image generation. However, due to the iterative nature of the diffusion process and its reliance on classifier-free guidance, inference times are slow. In this paper, we propose a new distillation approach for guided diffusion models in which an external lightweight guide model is trained while the original text-to-image model remains frozen.We show that our method reduces the inference computation of classifier-free guided latent-space diffusion models by almost half, and only requires 1% trainable parameters of the base model. Furthermore, once trained, our guide model can be applied to various fine-tuned, domain-specific versions of the base diffusion model without the need for additional training: this "plug-and-play" functionality drastically improves inference computation while maintaining the visual fidelity of generated images. Empirically, we show that our approach is able to produce visually appealing results and achieve a comparable FID score to the teacher with as few as 8 to 16 steps.

Vision-and-language navigation (VLN) enables the agent to navigate to a remote location following the natural language instruction in 3D environments. At each navigation step, the agent selects from possible candidate locations and then makes the move. For better navigation planning, the lookahead exploration strategy aims to effectively evaluate the agent's next action by accurately anticipating the future environment of candidate locations. To this end, some existing works predict RGB images for future environments, while this strategy suffers from image distortion and high computational cost. To address these issues, we propose the pre-trained hierarchical neural radiance representation model (HNR) to produce multi-level semantic features for future environments, which are more descriptive and efficientthan pixel-wise RGB reconstruction. Furthermore, with the predicted future environmental representations, our lookahead VLN model is able to construct the navigable future path tree and select the optimal path branch via efficient parallel evaluation. Extensive experiments on the VLN-CE datasets confirm the effectiveness of our proposed method.

This paper focuses on the high computational complexity in Large Language Models (LLMs), a significant challenge in both natural language processing (NLP) and multi-modal tasks. We propose Low-Rank Approximation for Sparse Attention (LoRA-Sparse), an innovative approach that strategically reduces this complexity. LoRA-Sparse introduces low-rank linear projection layers for sparse attention approximation. It utilizes an order-mimic training methodology, which is crucial for efficiently approximating the self-attention mechanism in LLMs. We empirically show that sparse attention not only reduces computational demands, but also enhances model performance in both NLP and multi-modal tasks. This surprisingly shows that redundant attention in LLMs might be non-beneficial. We extensively validate LoRA-Sparse through rigorous empirical studies in both (NLP) and multi-modal tasks, demonstrating its effectiveness and general applicability. Based on LLaMA and LLaVA models, our methods can reduce more than half of the self-attention computation with even better performance than full-attention baselines. Code will be made available.

Vision-Language Models (VLMs), such as CLIP, exhibit strong image-text comprehension abilities, facilitating advances in several downstream tasks such as zero-shot image classification, image-text retrieval, and text-to-image generation. However, the compositional reasoning abilities of existing VLMs remains subpar. The root of this limitation lies in the inadequate alignment between the images and captions in the pretraining datasets. Additionally, the current contrastive learning objective fails to focus on fine-grained grounding components like relations, actions, and attributes, resulting in "bag-of-words" representations. We introduce a simple and effective method to improve compositional reasoning in VLMs. Our method better leverages available datasets by refining and expanding the standard image-text contrastive learning framework. Our approach does not require specific annotations and does not incur extra parameters. When integrated with CLIP, our technique yields notable improvement over state-of-the-art baselines across five vision-language compositional benchmarks.

A fundamental characteristic common to both human vision and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, recent investigations find that most - if not all - our state-of-the-art vision-language models struggle at compositionality. They are unable to distinguish between images of "a girl in white facing a man in black" and "a girl in black facing a man in white". Moreover, prior work suggests that compositionality doesn't arise with scale: larger model sizes or training data don't help. This paper develops a new iterated training algorithm that incentivizes compositionality. We draw on decades of cognitive science research that identifies cultural transmission - the need to teach a new generation - as a necessary inductive prior that incentivizes humans to develop compositional languages. Specifically, we reframe vision-language contrastive learning as the Lewis Signaling Game between a vision agent and language agent, and operationalize cultural transmission by iteratively resetting one of the agent's weights during training. After every iteration, this training paradigm induces representations that become "easier to learn", a property of compositional languages: e.g. our model trained in CC3M and CC12M improves standard CLIP by 4.7%, 4.0% respectfully in the SugarCrepe benchmark.

Vision language models (VLMs) have experienced rapid advancements through the integration of large language models (LLMs) with image-text pairs, yet they struggle with detailed regional visual understanding due to limited spatial awareness of the vision encoder, and the use of coarse-grained training data that lacks detailed, region-specific captions.To address this, we introduce RegionGPT (short as RGPT), a novel framework designed for complex region-level captioning and understanding. RGPT enhances the spatial awareness of regional representation with simple yet effective modifications to existing visual encoders in VLMs. We further improve performance on tasks requiring a specific output scope by integrating task-guided instruction prompts during both training and inference phases, while maintaining the model's versatility for general-purpose tasks. Additionally, we develop an automated region caption data generation pipeline, enriching the training set with detailed region-level captions. We demonstrate that a universal RGPT model can be effectively applied and significantly enhancing performance across a range of region-level tasks, including but not limited to complex region descriptions, reasoning, object classification, and referring expressions comprehension.

Multimodal Large Language Models (MLLMs) have recently demonstrated impressive capabilities in multimodal understanding, reasoning, and interaction. However, existing MLLMs prevalently suffer from serious hallucination problems, generating text that is not factually grounded in associated images. The problem makes existing MLLMs untrustworthy and thus impractical in real-world (especially high-stakes) applications. To address the challenge, we present RLHF-V, which enhances MLLM trustworthiness via behavior alignment from fine-grained correctional human feedback. Specifically, RLHF-V collects human preference in the form of segment-level corrections on hallucinations, and performs dense direct preference optimization over the human feedback. Comprehensive experiments on five benchmarks in both automatic and human evaluation show that, RLHF-V can enable substantially more trustworthy MLLM behaviors with promising data and computation efficiency. Remarkably, using 1.4k annotated data samples, RLHF-V significantly reduces the hallucination rate of the base MLLM by 34.8%, outperforming the concurrent LLaVA-RLHF trained on 10k annotated data. The final model achieves state-of-the-art performance in trustworthiness among open-source MLLMs, and shows better robustness than GPT-4V in preventing hallucinations aroused from over-generalization. All the data, code and model weights will be released to facilitate future research.

In Multimodal Large Language Models (MLLMs), a visual projector plays a crucial role in bridging pre-trained vision encoders with LLMs, enabling profound visual understanding while harnessing the LLMs' robust capabilities. Despite the importance of the visual projector, it has been relatively less explored. In this study, we first identify two essential projector properties: (i) flexibility in managing the number of visual tokens, crucial for MLLMs' overall efficiency, and (ii) preservation of local context from visual features, vital for spatial understanding. Based on these findings, we propose a novel projector design that is both flexible and locality-enhanced, effectively satisfying the two desirable properties. Additionally, we present comprehensive strategies to effectively utilize multiple and multifaceted instruction datasets. Through extensive experiments, we examine the impact of individual design choices. Finally, our proposed MLLM, Honeybee, remarkably outperforms previous state-of-the-art methods across various benchmarks, including MME, MMBench, SEED-Bench, and LLaVA-Bench, achieving significantly higher efficiency. We will release the code and model publicly available.

Geometry Problem Solving (GPS) has drawn growing attention recently due to its application prospects in intelligent education field. However, existing methods are still inadequate to meet the needs of practical application, suffering from the following limitations: 1) explainability is not ensured which is essential in real teaching scenarios; 2) the small scale and incomplete annotation of existing datasets make it hard for model to learn geometric knowledge. To tackle the above problems, we propose a novel method called Explainable Geometry Problem Solving (E-GPS). E-GPS first parses the geometric diagram and problem text into unified formal language representations. Then, the answer and explainable reasoning and solving steps are obtained by a Top-Down Problem Solver (TD-PS), which innovatively solves the problem from the target and focuses on what is needed. To alleviate the data issues, a Bottom-Up Problem Generator (BU-PG) is devised to augment the data set with various well-annotated constructed geometry problems. It enables us to train an enhanced theorem predictor with a better grasp of theorem knowledge, which further improves the efficiency of TD-PS. Extensive experiments demonstrate that E-GPS maintains comparable solving performances with fewer steps and provides outstanding explainability.

Multi-modal Large Language Models (MLLMs) have shown remarkable capabilities in various multi-modal tasks. Nevertheless, their performance in fine-grained image understanding tasks is still limited. To address this issue, this paper proposes a new framework that aims to enhance the fine-grained image understanding abilities of MLLMs. Specifically, we present a new method for constructing the instruction tuning dataset at a low cost by leveraging annotations in existing datasets. A self-consistent bootstrapping method is also introduced to extend existing dense object annotations into high-quality referring-expression-bounding-box pairs. These methods enable the generation of high-quality instruction data which includes a wide range of fundamental abilities essential for fine-grained image perception. Moreover, we argue that the visual encoder should be tuned during instruction tuning to mitigate the gap between full image perception and fine-grained image perception. Experimental results demonstrate the superior performance of our method. For instance, our model exhibits a 5.2% accuracy improvement over Qwen-VL on GQA and surpasses the accuracy of Kosmos-2 by 24.7% on RefCOCO_val. We also attain the top rank on the leaderboard of MMBench. This promising performance is achieved by training on only publicly available data, making it easily reproducible. We will release the models, datasets, and codes for further research.

Image-language models with prompt learning have shown remarkable advances in numerous downstream vision tasks. Nevertheless, conventional prompt learning methods overfit the training distribution and lose the generalization ability on the test distributions. To improve the generalization across various distribution shifts, we propose any-shift prompting: a general probabilistic inference framework that considers the relationship between training and test distributions during prompt learning. We explicitly connect training and test distributions in the latent space by constructing training and test prompts in a hierarchical architecture. Within this framework, the test prompt exploits the distribution relationships to guide the generalization of the CLIP image-language model from training to any test distribution. To effectively encode the distribution information and their relationships, we further introduce a transformer inference network with a pseudo-shift training mechanism. The network generates the tailored test prompt with both training and test information in a feedforward pass, avoiding extra training costs at test time. Extensive experiments on twenty-three datasets demonstrate the effectiveness of any-shift prompting on the generalization over various distribution shifts.

Vision-Language (VL) models have gained significant research focus, enabling remarkable advances in multimodal reasoning. These architectures typically comprise a vision encoder, a Large Language Model (LLM), and a projection module that aligns visual features with the LLM's representation space. Despite their success, a critical limitation persists: the vision encoding process remains decoupled from user queries, often in the form of image-related questions. Consequently, the resulting visual features may not be optimally attuned to the query-specific elements of the image.To address this, we introduce QA-ViT, a Question Aware Vision Transformer approach for multimodal reasoning, which embeds question awareness directly within the vision encoder.This integration results in dynamic visual features focusing on relevant image aspects to the posed question.QA-ViT is model-agnostic and can be incorporated efficiently into any VL architecture.Extensive experiments demonstrate the effectiveness of applying our method to various multimodal architectures, leading to consistent improvement across diverse tasks and showcasing its potential for enhancing visual and scene-text understanding.

Large Vision-Language Models (LVLMs) have advanced considerably, intertwining visual recognition and language understanding to generate content that is not only coherent but also contextually attuned. Despite their success, LVLMs still suffer from the issue of object hallucinations, where models generate plausible yet incorrect outputs that include objects that do not exist in the images. To mitigate this issue, we introduce Visual Contrastive Decoding (VCD), a simple and training-free method that contrasts output distributions derived from original and distorted visual inputs. The proposed VCD effectively reduces the over-reliance on statistical bias and unimodal priors, two essential causes of object hallucinations. This adjustment ensures the generated content is closely grounded to visual inputs, resulting in contextually accurate outputs. Our experiments show that VCD, without either additional training or the usage of external tools, significantly mitigates the object hallucination issue across different LVLM families. Beyond mitigating object hallucinations, VCD also excels in general LVLM benchmarks, highlighting its wide-ranging applicability. Codes will be released.

Diffusion models are generative models with impressive text-to-image synthesis capabilities and have spurred a new wave of creative methods for classical machine learning tasks. However, the best way to harness the perceptual knowledge of these generative models for visual tasks is still an open question. Specifically, it is unclear how to use the prompting interface when applying diffusion backbones to vision tasks. We find that automatically generated captions can improve text-image alignment and significantly enhance a model's cross-attention maps, leading to better perceptual performance. Our approach improves upon the current state-of-the-art (SOTA) in diffusion-based semantic segmentation on ADE20K and the current overall SOTA in depth estimation on NYUv2. Furthermore, our method generalizes to the cross-domain setting. We use model personalization and caption modifications to align our model to the target domain and find improvements over unaligned baselines. Our cross-domain object detection model, trained on Pascal VOC, achieves SOTA results on Watercolor2K. Our cross-domain segmentation method, trained on Cityscapes, achieves SOTA results on Dark Zurich-val and Nighttime Driving.

There has been significant attention to the research on dense video captioning, which aims to automatically localize and caption all events within untrimmed video. Several studies introduce methods by designing dense video captioning as a multitasking problem of event localization and event captioning to consider inter-task relations. However, addressing both tasks using only visual input is challenging due to the lack of semantic content. In this study, we address this by proposing a novel framework inspired by the cognitive information processing of humans. Our model utilizes external memory to incorporate prior knowledge. The memory retrieval method is proposed with cross-modal video-to-text matching. To effectively incorporate retrieved text features, the versatile encoder and the decoder with visual and textual cross-attention modules are designed. Comparative experiments have been conducted to show the effectiveness of the proposed method on ActivityNet Captions and YouCook2 datasets. Experimental results show promising performance of our model without extensive pretraining from a large video dataset.

Recent dataset deduplication techniques have demonstrated that content-aware dataset pruning can dramatically reduce the cost of training Vision-Language Pretrained (VLP) models without significant performance losses compared to training on the original dataset. These results have been based on pruning commonly used image-caption datasets collected from the web -- datasets that are known to harbor harmful social biases that may then be codified in trained models. In this work, we evaluate how deduplication affects the prevalence of these biases in the resulting trained models and introduce an easy-to-implement modification to the recent SemDeDup algorithm that can reduce the negative effects that we observe. When examining CLIP-style models trained on deduplicated variants of LAION-400M, we find our proposed FairDeDup algorithm consistently leads to improved fairness metrics over SemDeDup on the FairFace and FACET datasets while maintaining zero-shot performance on CLIP benchmarks.

Grounding referred objects in 3D scenes is a burgeoning vision-language task pivotal for propelling Embodied AI, as it endeavors to connect the 3D physical world with free-form descriptions. Compared to the 2D counterparts, challenges posed by the variability of 3D visual grounding remain relatively unsolved in existing studies: 1) the underlying geometric and complex spatial relationships in 3D scene. 2) the inherent complexity of 3D grounded language. 3) the inconsistencies between text and geometric features. To tackle these issues, we propose G$^3$-LQ, a DEtection TRansformer-based model tailored for 3D visual grounding task. G$^3$-LQ explicitly models $\textbf{G}$eometric-aware visual representations and $\textbf{G}$enerates fine-$\textbf{G}$rained $\textbf{L}$anguage-guided object $\textbf{Q}$ueries in an overarching framework, which comprises two dedicated modules. Specifically, the Position Adaptive Geometric Exploring (PAGE) unearths underlying information of 3D objects in the geometric details and spatial relationships perspectives. The Fine-grained Language-guided Query Selection (Flan-QS) delves into syntactic structure of texts and generates object queries that exhibit higher relevance towards fine-grained text features. Finally, a pioneering Poincaré Semantic Alignment (PSA) loss establishes semantic-geometry consistencies by modeling non-linear vision-text feature mappings and aligning them on a hyperbolic prototype—Poincaré ball. Extensive experiments verify the superiority of our G$^3$-LQ method, trumping the state-of-the-arts by a considerable margin.

Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction or pre-defined categories to identify the target objects before executing visual recognition tasks. Such systems cannot actively reason and comprehend implicit user intention. In this work, we propose a new segmentation task --- reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction-mask data samples, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of multimodal Large Language Models (LLMs) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving complex reasoning and world knowledge. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation data samples results in further performance enhancement. Both quantitative and qualitative experiments show our method effectively unlocks new reasoning segmentation capabilities for multimodal LLMs. Code, models, and data are available at github.com/dvlab-research/LISA.

Despite being (pre)trained on a massive amount of data, state-of-the-art video-language alignment models are not robust to semantically-plausible contrastive changes in the video captions. Our work addresses this by identifying a broad spectrum of contrast misalignments, such as replacing entities, actions, and flipping event order, which alignment models should be robust against. To this end, we introduce the VideoCon, a video-language alignment dataset constructed by a large language model that generates plausible contrast video captions and explanations for differences between original and contrast video captions. Then, a generative video-language model is finetuned with VideoCon to assess video-language entailment and generate explanations. Our VideoCon-based alignment model significantly outperforms current models. It exhibits a $12$-point increase in AUC for the video-language alignment task on human-generated contrast captions. Finally, our model sets new state of the art zero-shot performance in temporally-extensive video-language tasks such as text-to-video retrieval (SSv2-Temporal) and video question answering (ATP-Hard). Moreover, our model shows superior performance on novel videos and human-crafted captions and explanations.

Recent studies have shown promising performance in open-vocabulary object detection (OVD) by utilizing pseudo labels (PLs) from pretrained vision and language models (VLMs). However, teacher-student self-training, a powerful and widely used paradigm to leverage PLs, is rarely explored for OVD. This work identifies two challenges of using self-training in OVD: noisy PLs from VLMs and frequent distribution changes of PLs. To address these challenges, we propose SAS-Det that tames self-training for OVD from two key perspectives. First, we present a split-and-fusion (SAF) head that splits a standard detection into an open-branch and a closed-branch. This design can reduce noisy supervision from pseudo boxes. Moreover, the two branches learn complementary knowledge from different training data, significantly enhancing performance when fused together. Second, in our view, unlike in closed-set tasks, the PL distributions in OVD are solely determined by the teacher model. We introduce a periodic update strategy to decrease the number of updates to the teacher, thereby decreasing the frequency of changes in PL distributions, which stabilizes the training process. Extensive experiments demonstrate SAS-Det is both efficient and effective. SAS-Det outperforms recent models of the same scale by a clear margin and achieves 37.4 AP50 and 29.1 APr on novel categories of the COCO and LVIS benchmarks, respectively. Code is available at https://github.com/xiaofeng94/SAS-Det.

Vision-language models (VLMs) have made significant strides in cross-modal understanding through large-scale paired datasets. However, in fashion domain, datasets often exhibit a disparity between the information conveyed in image and text. This issue stems from datasets containing multiple images of a single fashion item all paired with one text, leading to cases where some textual details are not visible in individual images. This mismatch, particularly when non-co-occurring elements are masked, undermines the training of conventional VLM objectives like Masked Language Modeling and Masked Image Modeling, thereby hindering the model’s ability to accurately align fine-grained visual and textual features. Addressing this problem, we propose Synchronized attentional Masking (SyncMask), which generate masks that pinpoint the image patches and word tokens where the information co-occur in both image and text. This synchronization is accomplished by harnessing cross-attentional features obtained from a momentum model, ensuring a precise alignment between the two modalities. Additionally, we enhance grouped batch sampling with semi-hard negatives, effectively mitigating false negative issues in Image-Text Matching and Image-Text Contrastive learning objectives within fashion datasets. Our experiments demonstrate the effectiveness of the proposed approach, outperforming existing methods in three downstream tasks.

In recent research, significant attention has been devoted to the open-vocabulary object detection task, aiming to generalize beyond the limited number of classes labeled during training and detect objects described by arbitrary category names at inference. Compared with conventional object detection, open vocabulary object detection largely extends the object detection categories.However, it relies on calculating the similarity between image regions and a set of arbitrary category names with a pretrained vision-and-language model.This implies that, despite its open-set nature, the task still needs the predefined object categories during the inference stage.This raises the question: What if we do not have exact knowledge of object categories during inference?In this paper, we call such a new setting as generative open-ended object detection, which is a more general and practical problem. To address it, we formulate object detection as a generative problem and propose a simple framework named GenerateU, which can detect dense objects and generate their names in a free-form way. Particularly, we employ Deformable DETR as a region proposal generator with a language model translating visual regions to object names.To assess the free-form object detection task, we introduce an evaluation method designed to quantitatively measure the performance of generative outcomes.Extensive experiments demonstrate strong zero-shot detection performance of our GenerateU. For example, on the LVIS dataset, our GenerateU achieves comparable results to the open-vocabulary object detection method GLIP, even though the category names are not seen by GenerateU during inference. Code will be released.

Diagram Question Answering (DQA) is a challenging task, requiring models to answer natural language questions based on visual diagram contexts. It serves as a crucial basis for academic tutoring, technical support, and more practical applications. DQA poses significant challenges, such as the demand for domain-specific knowledge and the scarcity of annotated data, which restrict the applicability of large-scale deep models. Previous approaches have explored external knowledge integration through pre-training, but these methods are costly and can be limited by domain disparities. While Large Language Models (LLMs) show promise in question-answering, there is still a gap in how to cooperate and interact with the diagram parsing process. In this paper, we introduce the Chain-of-Guiding Learning Model for Diagram Question Answering (CoG-DQA), a novel framework that effectively addresses DQA challenges. CoG-DQA leverages LLMs to guide diagram parsing tools (DPTs) through the guiding chains, enhancing the precision of diagram parsing while introducing rich background knowledge. Our experimental findings reveal that CoG-DQA surpasses all comparison models in various DQA scenarios, achieving an average accuracy enhancement exceeding 5\% and peaking at 11\% across four datasets. These results underscore CoG-DQA's capacity to advance the field of visual question answering and promote the integration of LLMs into specialized domains.

Multimodal Large Language Model (MLLMs) leverages Large Language Models as a cognitive framework for diverse visual-language tasks. Recent efforts have been made to equip MLLMs with visual perceiving and grounding capabilities.However, there still remains a gap in providing fine-grained pixel-level perceptions and extending interactions beyond text-specific inputs. In this work, we propose AnyRef, a general MLLM model that can generate pixel-wise object perceptions and natural language descriptions from multi-modality references, such as texts, boxes, images, or audios. This innovation empowers users with greater flexibility to engage with the model beyond textual and regional prompts, without modality-specific designs. Through our proposed refocusing mechanism, the generated grounding output is guided to focus more on the referenced object, implicitly incorporating additional pixel-level supervision. This simple modification utilizes attention scores generated during the inference of LLM, eliminating the need for extra computations while exhibiting performance enhancements in both grounding masks and referring expressions. With only publicly available training data, our model achieves state-of-the-art results across multiple benchmarks, including diverse modality referring segmentation and region-level referring expression generation.

Robotics agents often struggle to understand and follow the multi-modal prompts in complex manipulation scenes which are challenging to be sufficiently and accurately described by text alone. Moreover, for long-horizon manipulation tasks, the deviation from general instruction tends to accumulate if lack of intermediate guidance from high-level subgoals. For this, we consider can we generate subgoal images before act to enhance the instruction following in long-horizon manipulation with multi-modal prompts?Inspired by the great success of diffusion model in image generation tasks, we propose a novel hierarchical framework named as CoTDiffusion that incorporates diffusion model as a high-level planner to convert the general and multi-modal prompts into coherent visual subgoal plans, which further guide the low-level policy model before action execution.We design a semantic alignment module that can anchor the progress of generated keyframes along a coherent generation chain, unlocking the chain-of-thought reasoning ability of diffusion model. Additionally, we propose bi-directional generation and frame concat mechanism to further enhance the fidelity of generated subgoal images and the accuracy of instruction following. The experiments cover various robotics manipulation scenarios including visual reasoning, visual rearrange, and visual constraints. CoTDiffusion achieves outstanding performance gain compared to the baselines without explicit subgoal generation, which proves that a subgoal image is worth a thousand words of instruction.

Referring video object segmentation (RVOS) aims to segment the target instance referred by a given text expression in a video clip. The text expression normally contains sophisticated description of the instance's appearance, action, and relation with others. It is therefore rather difficult for a RVOS model to capture all these attributes correspondingly in the video; in fact, the model often favours more on the action- and relation-related visual attributes of the instance. This can end up with partial or even incorrect mask prediction of the target instance. We tackle this problem by taking a subject-centric short text expression from the original long text expression. The short one retains only the appearance-related information of the target instance so that we can use it to focus the model's attention on the instance's appearance. We let the model make joint predictions using both long and short text expressions; and insert a long-short cross-attention module to interact the joint features and a long-short predictions intersection loss to regulate the joint predictions. Besides the improvement on the linguistic part, we also introduce a forward-backward visual consistency loss, which utilizes optical flows to warp visual features between the annotated frames and their temporal neighbors for consistency. We build our method on top of two state of the art pipelines. Extensive experiments on A2D-Sentences, Refer-YouTube-VOS, JHMDB-Sentences and Refer-DAVIS17 show impressive improvements of our method. Code is available here (https://github.com/LinfengYuan1997/LoSh).

Characters are an important aspect of any storyline and identifying and including them in descriptions is necessary for story understanding. While previous work has largely ignored identity and generated captions with someone (anonymized names), recent work formulates id-aware captioning as a fill-in-the-blanks (FITB) task, where, given a caption with blanks, the goal is to predict person id labels. However, to predict captions with ids, a two-stage approach is required: first predict captions with someone, then fill in identities. In this work, we present a new single stage approach that can seamlessly switch between id-aware caption generation or FITB when given a caption with blanks. Our model, Movie-Identity Captioner (MICap), uses a shared auto-regressive decoder that benefits from training with FITB and full-caption generation objectives, while the encoder can benefit from or disregard captions with blanks as input. Another challenge with id-aware captioning is the lack of a metric to capture subtle differences between person ids. To this end, we introduce iSPICE, a caption evaluation metric that focuses on identity tuples created through intermediate scene graphs. We evaluate MICap on Large-Scale Movie Description Challenge (LSMDC), where we show a 4.2% improvement in FITB accuracy, and a 1-2% bump in classic captioning metrics.

Large multimodal models demonstrate remarkable generalist ability to perform diverse multimodal tasks in a zero-shot manner. Large-scale web-based image-text pairs contribute fundamentally to this success, but suffer from excessive noise. Recent studies use alternative captions synthesized by captioning models and have achieved notable benchmark performance. However, our experiments reveal significant Scalability Deficiency and World Knowledge Loss issues in models trained with synthetic captions, which have been largely obscured by their initial benchmark success. Upon closer examination, we identify the root cause as the overly-simplified language structure and lack of knowledge details in existing synthetic captions. To provide higher-quality and more scalable multimodal pretraining data, we propose CapsFusion, an advanced framework that leverages large language models to consolidate and refine information from both web-based image-text pairs and synthetic captions. Extensive experiments show that CapsFusion captions exhibit remarkable all-round superiority over existing captions in terms of model performance (e.g., 18.8 and 18.3 improvements in CIDEr score on COCO and NoCaps), sample efficiency (requiring 11-16 times less computation than baselines), world knowledge depth, and scalability. These effectiveness, efficiency and scalability advantages position CapsFusion as a promising candidate for future scaling of LMM training.

Existing automatic captioning methods for visual content face challenges such as lack of detail, content hallucination, and poor instruction following. In this work, we propose VisualFactChecker (VFC), a flexible training-free pipeline that generates high-fidelity and detailed captions for both 2D images and 3D objects. VFC consists of three steps: 1) proposal, where image-to-text captioning models propose multiple initial captions; 2) verification, where a large language model (LLM) utilizes tools such as object detection and VQA models to fact-check proposed captions; 3) captioning, where an LLM generates the final caption by summarizing caption proposals and the fact check verification results. In this step, VFC can flexibly generate captions in various styles following complex instructions. We conduct comprehensive captioning evaluations using four metrics: 1) CLIP-Score for image-text similarity; 2) CLIP-Image-Score for measuring the image-image similarity between the original and the reconstructed image generated by a text-to-image model using the caption. 3) human study on Amazon Mechanical Turk; 4) GPT-4V for fine-grained evaluation. Evaluation results show that VFC outperforms state-of-the-art open-sourced captioning methods for 2D images on the COCO dataset and 3D assets on the Objaverse dataset. Our study demonstrates that by combining open-source models into a pipeline, we can attain captioning capability comparable to proprietary models such as GPT-4V, despite being over 10$\times$ smaller in model size.

In this paper, we propose VidLA, an approach for video-language alignment at scale. There are two major limitations of previous video-language alignment approaches. First, they do not capture both short-range and long-range temporal dependencies, and typically employ complex hierarchical deep network architectures that are hard to integrate with existing pretrained image-text foundation models. To effectively address this limitation, we instead keep the network architecture simple and use a set of data tokens that operate at different temporal resolutions in a hierarchical manner, accounting for the temporally hierarchical nature of videos. By employing a simple two-tower architecture, we are able to initialize our video-language model with pretrained image-text foundation models, thereby boosting the final performance. Second, existing video-language alignment works struggle due to the lack of semantically aligned large-scale training data. To overcome it, we leverage recent LLMs to curate the largest video-language dataset to-date with better visual grounding. Furthermore, unlike existing video-text datasets which only contains short clips, our dataset is enriched with video clips of varying durations to aid our temporally hierarchical data tokens in extracting better representations at varying temporal scales. Overall, empirical results show that our proposed approach surpasses state-of-the-art methods on multiple retrieval benchmarks, especially on longer videos, and performs competitively on classification benchmarks.

Referring expressions for visual objects often include descriptions of relative spatial arrangements to other objects -- e.g. ``to the right of'' -- that depend on the point of view of the speaker. In 2D referring expression tasks, this viewpoint is captured unambiguously in the image. However, grounding expressions with such spatial language in 3D without viewpoint annotations can be ambiguous. In this paper, we investigate the significance of viewpoint information in 3D visual grounding -- introducing a model that explicitly predicts the speaker's viewpoint based on the referring expression and scene. We pretrain this model on a synthetically generated dataset that provides viewpoint annotations and then finetune on 3D referring expression datasets. Further, we introduce an auxiliary uniform object representation loss to encourage viewpoint invariance in learned object representations. We find that our proposed ViewPoint Prediction Network (VPP-Net) achieves state-of-the-art performance on ScanRefer, SR3D, and NR3D -- improving Accuracy@0.25IoU by 1.06\%, 0.60\%, and 4.80\% respectively compared to prior work.

Multiple clustering has gained significant attention in recent years due to its potential to reveal multiple hidden structures of data from different perspectives. The advent of deep multiple clustering techniques has notably advanced the performance by uncovering complex patterns and relationships within large datasets. However, a major challenge arises as users often do not need all the clusterings that algorithms generate, and figuring out the one needed requires a substantial understanding of each clustering result. Traditionally, aligning a user's brief keyword of interest with the corresponding vision components was challenging, but the emergence of multi-modal and large language models (LLMs) has begun to bridge this gap. In response, given unlabeled target visual data, we propose Multi-Map, a novel method employing a multi-modal proxy learning process. It leverages CLIP encoders to extract coherent text and image embeddings, with GPT-4 integrating users' interests to formulate effective textual contexts. Moreover, reference word constraint and concept-level constraint are designed to learn the optimal text proxy according to the user’s interest. Multi-Map not only adeptly captures a user's interest via a keyword but also facilitates identifying relevant clusterings. Our extensive experiments show that Multi-Map consistently outperforms state-of-the-art methods in all benchmark multi-clustering vision tasks. Our code is available at https://github.com/Alexander-Yao/Multi-MaP.

The ability of large language models (LLMs) to process visual inputs has given rise to general-purpose vision systems, unifying various vision-language (VL) tasks by instruction tuning. However, due to the enormous diversity in input-output formats in the vision domain, existing general-purpose models fail to successfully integrate segmentation and multi-image inputs with coarse-level tasks into a single framework. In this work, we introduce VistaLLM, a powerful visual system that addresses coarse- and fine-grained VL tasks over single and multiple input images using a unified framework. VistaLLM utilizes an instruction-guided image tokenizer that filters global embeddings using task descriptions to extract compressed and refined features from numerous images. Moreover, VistaLLM employs a gradient-aware adaptive sampling technique to represent binary segmentation masks as sequences, significantly improving over previously used uniform sampling. To bolster the desired capability of VistaLLM, we curate CoinIt, a comprehensive coarse-to-fine instruction tuning dataset with $6.8$M samples. We also address the lack of multi-image grounding datasets by introducing a novel task, AttCoSeg (Attribute-level Co-Segmentation), which boosts the model's reasoning and grounding capability over multiple input images. Extensive experiments on a wide range of V- and VL tasks demonstrate the effectiveness of VistaLLM by achieving consistent state-of-the-art performance over strong baselines across all downstream tasks. Code and data for training VistaLLM will be publicly released.

Existing methods to fine-tune LLMs, like Adapter, Prefix-tuning, and LoRA, which introduce extra modules or additional input sequences to inject new skills or knowledge, may compromise the innate abilities of LLMs.In this paper, we propose LLaMA-Excitor, a lightweight method that stimulates the LLMs' potential to better follow instructions by gradually paying more attention to worthwhile information.Specifically, the LLaMA-Excitor does not directly change the intermediate hidden state during the self-attention calculation of the transformer structure.We designed the Excitor block as a bypass module for the similarity score computation in LLMs' self-attention to reconstruct keys and change the importance of values by learnable prompts. LLaMA-Excitor ensures a self-adaptive allocation of additional attention to input instructions, thus effectively preserving LLMs' pre-trained knowledge when fine-tuning LLMs on low-quality instruction-following datasets. Furthermore, we unify the modeling of multi-modal tuning and language-only tuning, extending LLaMA-Excitor to a powerful visual instruction follower without the need for complex multi-modal alignment. Our proposed approach is evaluated in language-only and multi-modal tuning experimental scenarios. Notably, LLaMA-Excitor is the only method that maintains basic capabilities while achieving a significant improvement (+6\%) on the MMLU benchmark.In the visual instruction tuning, we achieve a new state-of-the-art image captioning performance of 157.5 CIDEr on MSCOCO, and a comparable performance (88.39\%) on ScienceQA to cutting-edge models with more parameters and extensive vision-language pertaining.

Zero-shot image captioning (IC) without well-paired image-text data can be divided into two categories: training-free and text-only-training. Generally, these two types of methods realize zero-shot IC by integrating pre-trained vision-language models like CLIP for image-text similarity evaluation and a pre-trained language model (LM) for caption generation. The main difference between them is whether to use a textual corpus to train the LM. Despite achieving attractive performance with respect to some metrics, existing methods often exhibit common drawbacks. Training-free methods tend to produce hallucinations, while text-only-training methods often lose generalization capability. To advance the field, this paper proposes a novel Memory-Augmented zero-shot image Captioning framework (MeaCap). Specifically, equipped with textual memory, we introduce a retrieve-then-filter module to extract key concepts highly related to the image. By deploying our proposed memory-augmented visual-related fusion score in a keywords-to-sentence LM, MeaCap can generate concept-centered captions that maintain high consistency with the image, reducing hallucinations and incorporating more world knowledge. The MeaCap framework achieves state-of-the-art performance across a series of zero-shot IC settings. The code is provided in the Supplement for further exploration.

Recognizing continuous changes offers valuable insights into past historical events, supports current trend analysis, and facilitates future planning. This knowledge is crucial for a variety of fields, such as meteorology and agriculture, environmental science, urban planning and construction, tourism, and cultural preservation. Currently, available datasets in the field of scene change understanding primarily concentrate on two main tasks: the detection of changed regions within a scene and the linguistic description of the change content. Existing datasets focus on recognizing discrete changes, such as adding or deleting an object from two images, and largely rely on artificially generated images. Consequently, the existing change understanding methods primarily focus on identifying distinct object differences, overlooking the importance of continuous, gradual changes occurring over extended time intervals. To address the above issues, we propose a novel benchmark dataset, STVchrono, targeting the localization and description of long-term continuous changes in real-world scenes. The dataset consists of 71,900 photographs from Google Street View API taken over an 18-year span across 50 cities all over the world. Our STVchrono dataset is designed to support change recognition and description in both image pairs and extended image sequences, while also enabling the segmentation of changed regions. We conduct experiments to evaluate state-of-the-art methods on change description and segmentation, as well as multimodal Large Language Models for describing changes. Our findings reveal that even the most advanced methods lag human performance, emphasizing the need to adapt them to continuously changing real-world scenarios. We hope that our benchmark dataset will further facilitate the research of temporal change recognition in a dynamic world.

In this paper, we present a novel paradigm to enhance the ability of object detector, e.g., expanding categories or improving detection performance, by training on syn- thetic dataset generated from diffusion models. Specifically, we integrate an instance-level grounding head into a pre- trained, generative diffusion model, to augment it with the ability of localising instances in the generated images. The grounding head is trained to align the text embedding of category names with the regional visual feature of the diffusion model, using supervision from an off-the-shelf object detector, and a novel self-training scheme on (novel) categories not covered by the detector. We conduct thorough experiments to show that, this enhanced version of diffusion model, termed as InstaGen, can serve as a data synthesizer, to enhance object detectors by training on its generated samples, demonstrating superior performance over existing state-of-the-art methods in open-vocabulary (+4.5 AP) and data-sparse (+1.2 ∼ 5.2 AP) scenarios.

3D visual grounding involves matching natural language descriptions with their corresponding objects in 3D spaces. Existing methods often face challenges with accuracy in object recognition and struggle in interpreting complex linguistic queries, particularly with descriptions that involve multiple anchors or are view-dependent. In response, we present the MiKASA (Multi-Key-Anchor Scene-Aware) Transformer. Our novel model integrates a self-attention-based scene-aware object encoder and an original multi-key-anchor technique, enhancing object recognition accuracy and the understanding of spatial relationships. Furthermore, MiKASA improves the explainability of decision-making, facilitating error diagnosis. Our model outperforms state-of-the-art models in the Referit3D challenge by a large margin, achieving higher accuracy across all categories for both Sr3d and Nr3d datasets, with overall improvements of 8.2% and 8.4%, respectively. This is especially pronounced in categories requiring viewpoint-dependent descriptions. We will make our code available upon the paper's publication.

Pre-trained vision-language models (VLMs) have achieved high performance on various downstream tasks, which have been widely used for visual grounding tasks in a weakly supervised manner. However, despite the performance gains contributed by large vision and language pre-training, we find that state-of-the-art VLMs struggle with compositional reasoning on grounding tasks. To demonstrate this, we propose Attribute, Relation, and Priority Grounding (ARPGrounding) benchmark to test VLMs' compositional reasoning ability on visual grounding tasks. ARPGrounding contains 11,425 samples and evaluates the compositional understanding of VLMs in three dimensions: 1) attribute, denoting comprehension of objects' properties, 2) relation, indicating an understanding of relation between objects, 3) priority, reflecting an awareness of the part of speech associated with nouns. Using the ARPGrounding benchmark, we evaluate several mainstream VLMs. We empirically find that these models perform quite well on conventional visual grounding datasets, achieving performance comparable to or surpassing state-of-the-art methods. However, they show strong deficiencies in compositional reasoning, as evidenced by their inability to establish links between objects and their associated attributes, a limited grasp of relational understanding, and insensitivity towards the prioritization of objects. Furthermore, we propose a composition-aware fine-tuning pipeline, demonstrating the potential to leverage cost-effective image-text annotations for enhancing the compositional understanding of VLMs in grounding tasks.

Inspired by the success of general-purpose models in NLP, recent studies attempt to unify different vision tasks in the same sequence format and employ autoregressive Transformers for sequence prediction. They apply uni-directional attention to capture sequential dependencies and generate task sequences recursively. However, such autoregressive Transformers may not fit vision tasks well, as vision task sequences usually lack the sequential dependencies typically observed in natural languages. In this work, we design Masked AutoDecoder~(MAD), an effective multi-task vision generalist. MAD consists of two core designs. First, we develop a parallel decoding framework that introduces bi-directional attention to capture contextual dependencies comprehensively and decode vision task sequences in parallel. Second, we design a masked sequence modeling approach that learns rich task contexts by masking and reconstructing task sequences. In this way, MAD handles all the tasks by a single network branch and a simple cross-entropy loss with minimal task-specific designs. Extensive experiments demonstrate the great potential of MAD as a new paradigm for unifying various vision tasks. MAD achieves superior performance and inference efficiency compared to autoregressive counterparts while obtaining competitive accuracy with task-specific models. Code will be released.

Test-time adaptation with pre-trained vision-language models has attracted increasing attention for tackling distribution shifts during the test time. Though prior studies have achieved very promising performance, they involve intensive computation which is severely unaligned with test-time adaptation. We design TDA, a training-free dynamic adapter that enables effective and efficient test-time adaptation with vision-language models. TDA works with a lightweight key-value cache that maintains a dynamic queue with few-shot pseudo labels as values and the corresponding test-sample features as keys. Leveraging the key-value cache, TDA allows adapting to test data gradually via progressive pseudo label refinement which is super-efficient without incurring any backpropagation. In addition, we introduce negative pseudo labeling that alleviates the adverse impact of pseudo label noises by assigning pseudo labels to certain negative classes when the model is uncertain about its pseudo label predictions. Extensive experiments over two benchmarks demonstrate TDA’s superior effectiveness and efficiency as compared with the state-of-the-art. The code has been released in https://kdiaaa.github.io/tda/.

Despite noise and caption quality having been acknowledged as important factors impacting vision-language contrastive pre-training, in this paper, we show that the full potential of improving the training process by addressing such issues is yet to be realized. Specifically, we firstly study and analyze two issues affecting training: incorrect assignment of negative pairs, and low caption quality and diversity. Then, we devise effective solutions for addressing both problems, which essentially require training with multiple true positive pairs. Finally, we propose training with sigmoid loss to address such a requirement. We show very large gains over the current state-of-the-art for both image recognition ($\sim +6\%$ on average over 11 datasets) and image retrieval ($\sim +19\%$ on Flickr30k and $\sim +15\%$ on MSCOCO).

Current approaches for 3D scene graph prediction rely on labeled datasets to train models for a fixed set of known object classes and relationship categories. We present Open3DSG, an alternative approach to learn 3D scene graph prediction in an open world without requiring labeled scene graph data. We co-embed the features from a 3D scene graph prediction backbone with the feature space of powerful open world 2D vision language foundation models. This enables us to predict 3D scene graphs from 3D point clouds in a zero-shot manner by querying object classes from an open vocabulary and predicting the inter-object relationships from a grounded LLM with scene graph features and queried object classes as context. Open3DSG is the first 3D point cloud method to predict not only explicit open-vocabulary object classes, but also open-set relationships that are not limited to a predefined label set, making it possible to express rare as well as specific objects and relationships in the predicted 3D scene graph. Our experiments show that Open3DSG is effective at predicting arbitrary object classes as well as their complex inter-object relationships describing spatial, supportive, semantic and comparative relationships.

Automatic radiology report generation can provide substantial advantages to clinical physicians by effectively reducing their workload and improving efficiency. Despite the promising potential of current methods, challenges persist in effectively extracting and preventing degradation of prominent features, as well as enhancing attention on pivotal regions. In this paper, we propose an Instance-level Expert Knowledge and Aggregate Discriminative Attention framework (EKAGen) for Radiology Report Generation. We convert expert reports into an embedding space and generate comprehensive representations for each disease, which serve as Preliminary Knowledge Support (PKS). To prevent feature disruption, we select the representations in the embedding space with the smallest distances to PKS as Rectified Knowledge Support (RKS). Then, EKAGen diagnoses the diseases and retrieves knowledge from RKS, creating Instance-level Expert Knowledge (IEK) for each query image, boosting generation. Additionally, we introduce Aggregate Discriminative Attention Map (ADM), which utilize weak supervision to generate maps of pivotal regions used to emphasize discriminative regions. For training, we propose a Global Information Self-Distillation (GID) strategy, using an iteratively optimized model to distill global knowledge into EKAGen. Extensive experiments and analyses on IU X-Ray and MIMIC-CXR datasets demonstrate that EKAGen outperforms previous state-of-the-art methods.

Large multi-modal models (LMMs) exhibit remarkable performance across numerous tasks. However, generalist LMMs often suffer from performance degradation when tuned over a large collection of tasks. Recent research suggests that Mixture of Experts (MoE) architectures are useful for instruction tuning, but for LMMs of parameter size around O(50-100B), the prohibitive cost of replicating and storing the expert models severely limits the number of experts we can use.We propose Omni-SMoLA, an architecture that uses the Soft MoE approach to (softly) mix many multimodal low rank experts, and avoids introducing a significant number of new parameters compared to conventional MoE models. The core intuition here is that the large model provides a foundational backbone, while different lightweight experts residually learn specialized knowledge, either per-modality or multimodally. Extensive experiments demonstrate that the SMoLA approach helps improve the generalist performance across a broad range of generative vision-and-language tasks, achieving new SoTA generalist performance that often matches or outperforms single specialized LMM baselines, as well as new SoTA specialist performance.

Recent advancements in Vision-Language Models (VLMs) have marked a significant leap in bridging the gap between computer vision and natural language processing. However, traditional VLMs, trained through contrastive learning on limited and noisy image-text pairs, often lack the spatial and linguistic understanding to generalize well to dense vision tasks or less common languages. Our approach, SF-CLIP, circumvents this issue by implicitly building on the solid visual and language understanding of foundational models trained on vast amounts of unimodal data. SF-CLIP integrates contrastive image-text pretraining with a masked knowledge distillation from large foundational text and vision models. This methodology guides our VLM in developing robust text and image representations.As a result, SF-CLIP shows exceptional zero-shot classification accuracy and enhanced image and text retrieval capabilities, setting a new state of the art for ViT-B/16 trained on YFCC15M and CC12M. Moreover, the dense per-patch supervision enhances our zero-shot and linear probe performance in semantic segmentation tasks. A remarkable aspect of our model is its multilingual proficiency, evidenced by strong retrieval results in multiple languages despite being trained predominantly on English data. We achieve all of these improvements without sacrificing the training efficiency through our selective application of masked distillation and the inheritance of teacher word embeddings.

Most multimodal large language models (MLLMs) learn language-to-object grounding through causal language modeling where grounded objects are captured by bounding boxes as sequences of location tokens. This paradigm lacks pixel-level representations that are important for fine-grained visual understanding and diagnosis. In this work, we introduce GROUNDHOG, an MLLM developed by Grounding Large Language Models to holistic Segmentation. GROUNDHOG incorporates a masked feature extractor and converts extracted features into visual entity tokens for the MLLM backbone, which then connects groundable phrases to unified grounding masks by retrieving and merging the entity masks. To train GROUNDHOG, we carefully curated a grounded visual instruction tuning dataset - Multi-Modal Multi-Grained Grounding (M3G2) - by harvesting a collection of segmentation-grounded datasets with rich annotations. Our experimental results show that GROUNDHOG achieves superior performance on various language grounding tasks without task-specific fine-tuning. GROUNDHOG demonstrates better grounding towards complex forms of visual input and provides easy-to-understand diagnosis in failure cases.

Solving complex visual tasks such as ``Who invented the musical instrument on the right?'' involves a composition of skills: understanding space, recognizing instruments, and also retrieving prior knowledge. Recent work shows promise by decomposing such tasks using a large language model (LLM) into an executable program that invokes specialized vision models. However, generated programs are error-prone: they omit necessary steps, include spurious ones, and are unable to recover when the specialized models give incorrect outputs.Moreover, they require loading multiple models, incurring high latency and computation costs. We propose Visual Program Distillation (VPD), an instruction tuning framework that produces a vision-language model (VLM) capable of solving complex visual tasks with a single forward pass. VPD distills the reasoning ability of LLMs by using them to sample multiple candidate programs, which are then executed and verified to identify a correct one. It translates each correct program into a language description of the reasoning steps, which are then distilled into a VLM. Extensive experiments show that VPD improves the VLM's ability to count, understand spatial relations, and reason compositionally. Our VPD-trained PaLI-X outperforms all prior VLMs, achieving state-of-the-art performance across complex vision tasks, including MMBench, OK-VQA, A-OKVQA, TallyQA, POPE, and Hateful Memes. An evaluation with human annotators also confirms that VPD improves model response factuality and consistency. Finally, experiments on content moderation demonstrate that VPD is also helpful for adaptation to real-world applications with limited data.

We present DRESS, a large vision language model (LVLM) that innovatively exploits Natural Language feedback (NLF) from Large Language Models to enhance its alignment and interactions by addressing two key limitations in the state-of-the-art LVLMs. First, prior LVLMs generally rely only on the instruction finetuning stage to enhance alignment with human preferences. Without incorporating extra feedback, they are still prone to generate unhelpful, hallucinated, or harmful responses. Second, while the visual instruction tuning data is generally structured in a multi-turn dialogue format, the connections and dependencies among consecutive conversational turns are weak. This reduces the capacity for effective multi-turn interactions. To tackle these, we propose a novel categorization of the NLF into two key types: critique and refinement. The critique NLF identifies the strengths and weaknesses of the responses and is used to align the LVLMs with human preferences. The refinement NLF offers concrete suggestions for improvement and is adopted to improve the interaction ability of the LVLMs-- which focuses on LVLMs' ability to refine responses by incorporating feedback in multi-turn interactions. To address the non-differentiable nature of NLF, we generalize conditional reinforcement learning for training. Our experimental results demonstrate that DRESS can generate more helpful (9.76%), honest (11.52%), and harmless (21.03%) responses, and more effectively learn from feedback during multi-turn interactions compared to SOTA LVLMs.

Affordance detection in 3D data is key for bridging perception and action in robots. Existing efforts mostly focus on the visual aspect and overlook the interaction with linguistic instructions, thus limiting their generalization to new objects and integration with large language models (LLMs) which are excellent instruction interpreters. With this regard, we propose a novel task, Language-guided Affordance Segmentation on 3D Object (LASO). LASO challenges a model to segment a 3D object's part relevant to a given affordance question. To facilitate the task, we contribute a dataset comprising 19,751 point-question pairs, covering 8434 object shapes and 870 expert-crafted questions. As a pioneer solution, we further propose PointRefer, which highlights an adaptive fusion module to identify target affordance regions at different scales. To ensure a text-aware segmentation, we adopt a set of affordance queries conditioned on linguistic cues to generate dynamic kernels. These kernels are further used to convoluted with point features and generate a segmentation mask. Comprehensive experiments and analyses validate PointRefer's effectiveness. With these efforts, We hope that LASO can steer the direction of 3D affordance, guiding it towards enhanced integration with the evolving capabilities of LLMs. Code is available at https://anonymous.4open.science/r/LASO/.

Unsupervised visual grounding methods alleviate the issue of expensive manual annotation of image-query pairs by generating pseudo-queries. However, existing methods are prone to confusing the spatial relationships between objects and rely on designing complex prompt modules to generate query texts, which severely impedes the ability to generate accurate and comprehensive queries due to ambiguous spatial relationships and manually-defined fixed templates. To tackle these challenges, we propose a omni-directional language query generation approach for unsupervised visual grounding named Omni-Q. Specifically, we develop a 3D spatial relation module to extend the 2D spatial representation to 3D, thereby utilizing 3D location information to accurately determine the spatial position among objects. Besides, we introduce a spatial graph module, leveraging the power of graph structures to establish accurate and diverse object relationships and thus enhancing the flexibility of query generation.Extensive experiments on five public benchmark datasets demonstrate that our method significantly outperforms existing state-of-the-art unsupervised methods by up to 16.17%. In addition, when applied in the supervised setting, our method can freely save up to 60% human annotations without a loss of performance.

Large language models (LLMs) have shown remarkable text understanding capabilities, which have been extended as Video LLMs to handle video data for comprehending visual details. However, existing Video LLMs can only provide a coarse description of the entire video, failing to capture the precise start and end time boundary of specific events. In this paper, we solve this issue via proposing VTimeLLM, a novel Video LLM designed for fine-grained video moment understanding and reasoning with respect to time boundary. Specifically, our VTimeLLM adopts a boundary-aware three-stage training strategy, which respectively utilizes image-text pairs for feature alignment, multiple-event videos to increase temporal-boundary awareness, and high-quality video-instruction tuning to further improve temporal understanding ability as well as align with human intents. Extensive experiments demonstrate that in fine-grained time-related comprehension tasks for videos such as Temporal Video Grounding and Dense Video Captioning, VTimeLLM significantly outperforms existing Video LLMs. Besides, benefits from the fine-grained temporal understanding of the videos further enable VTimeLLM to beat existing Video LLMs in video dialogue benchmark, showing its superior cross-modal understanding and reasoning abilities.

People are spending an enormous amount of time on digital devices through graphical user interfaces (GUIs), e.g., computer or smartphone screens. Large language models (LLMs) such as ChatGPT can assist people in tasks like writing emails, but struggle to understand and interact with GUIs, thus limiting their potential to increase automation levels. In this paper, we introduce CogAgent, an 18-billion-parameter visual language model (VLM) specializing in GUI understanding and navigation. By utilizing both low-resolution and high-resolution image encoders, CogAgent supports input at a resolution of $1120\times 1120$, enabling it to recognize tiny page elements and text. As a generalist visual language model, CogAgent achieves the state of the art on five text-rich and four general VQA benchmarks, including VQAv2, OK-VQA, Text-VQA, ST-VQA, ChartQA, infoVQA, DocVQA, MM-Vet, and POPE. CogAgent, using only screenshots as input, outperforms LLM-based methods that consume extracted HTML text on both PC and Android GUI navigation tasks---Mind2Web and AITW, advancing the state of the art. The model and codes are available at https://github.com/THUDM/CogVLM.

Vision-language models (VLMs) have recently shown promising results in traditional downstream tasks.Evaluation studies have emerged to assess their abilities, with the majority focusing on the third-person perspective, and only a few addressing specific tasks from the first-person perspective.However, the capability of VLMs to "think" from a first-person perspective, a crucial attribute for advancing autonomous agents and robotics, remains largely unexplored. To bridge this research gap, we introduce EgoThink, a novel visual question-answering benchmark that encompasses six core capabilities with twelve detailed dimensions.The benchmark is constructed using selected clips from egocentric videos, with manually annotated question-answer pairs containing first-person information. To comprehensively assess VLMs, we evaluate twenty-one popular VLMs on EgoThink. Moreover, given the open-ended format of the answers, we use GPT-4 as the automatic judge to compute single-answer grading.Experimental results indicate that although GPT-4V leads in numerous dimensions, all evaluated VLMs still possess considerable potential for improvement in first-person perspective tasks.Meanwhile, enlarging the number of trainable parameters has the most significant impact on model performance on EgoThink.In conclusion, EgoThink serves as a valuable addition to existing evaluation benchmarks for VLMs, providing an indispensable resource for future research in the realm of embodied artificial intelligence and robotics.

Generative Vision-Language Models (VLMs) are prone to generate plausible-sounding textual answers which, however, are not always grounded in the input image. We investigate this phenomenon, usually referred to as "hallucination", and show that it stems from an excessive reliance on the language prior. In particular, we show that as more tokens are generated the reliance on the visual prompt decreases and this behavior strongly correlates with the emergence of hallucinations. To reduce hallucinations, we introduce Multi-Modal Mutual-Information Decoding (M3ID), a new sampling method for prompt amplification. M3ID amplifies the influence of the reference image over the language prior, hence favoring the generation of tokens with higher mutual information with the visual prompt. M3ID can be applied to any pre-trained autoregressive VLM at inference time without necessitating further training and with minimal computational overhead. If training is an option, we show that M3ID can be paired with Direct Preference Optimization (DPO) to improve the model's reliance on the prompt image without requiring any labels. Our empirical findings show that our algorithms maintain the fluency and linguistic capabilities of pre-trained VLMs while reducing hallucinations by mitigating visually ungrounded answers. Specifically, for the LLaVA 13B model, M3ID and M3ID+DPO reduce the percentage of hallucinated objects in captioning tasks by 25% and 28%, respectively, and improve the accuracy on VQA benchmarks such as POPE by 21% and 24%.

This work proposes TimeChat, a time-sensitive multimodal large language model specifically designed for long video understanding. Our model incorporates two key architectural contributions: (1) a timestamp-aware frame encoder that binds visual content with the timestamp of each frame, and (2) a sliding video Q-Former that produces a video token sequence of varying lengths to accommodate videos of various durations. Additionally, we construct an instruction-tuning dataset, encompassing 6 tasks and a total of 125K instances, to further enhance TimeChat's instruction-following performance. Experiment results across various video understanding tasks, such as dense captioning, temporal grounding, and highlight detection, demonstrate TimeChat's strong zero-shot temporal localization and reasoning capabilities. For example, it achieves +9.2 F1 score and +2.8 CIDEr on YouCook2, +5.8 HIT@1 on QVHighlights, and +27.5 R@1 (IoU=0.5) on Charades, compared to state-of-the-art video large language models, holding the potential to serve as a versatile video assistant for long-form video comprehension tasks and satisfy realistic user requirements.

Automatic radiology report generation using deep learning models has been recently explored and found promising. Neural decoders are commonly used for the report generation, where irrelevant and unfaithful contents are unavoidable. The retrieval-based approach alleviates the limitation by identifying reports which are relevant to the input to assist the generation. To achieve clinically accurate report retrieval, we make reference to clinicians' diagnostic steps of examining a radiology image where anatomical and diagnostic details are typically focused, and propose a novel hierarchical visual concept representation called anatomy-aware hierarchical vision encoding (AHIVE). To learn AHIVE, we first derive a methodology to extract hierarchical diagnostic descriptions from radiology reports and develop a CLIP-based framework for the model training. Also, the hierarchical architecture of AHIVE is designed to support interactive report retrieval so that report revision made at one layer can be propagated to the subsequent ones to trigger other necessary revisions. We conduct extensive experiments and show that AHIVE can outperform the SOTA vision-language retrieval methods in terms of clinical accuracy by a large margin. We provide also a case study to illustrate how it enables interactive report retrieval.

Aligned text-image encoders such as CLIP have become the de-facto model for vision-language tasks. Furthermore, modality-specific encoders achieve impressive performances in their respective domains. This raises a central question: does an alignment exist between uni-modal vision and language encoders since they fundamentally represent the same physical world? Analyzing the latent spaces structure of vision and language models on image-caption benchmarks using the Centered Kernel Alignment (CKA), we find that the representation spaces of unaligned and aligned encoders are semantically similar. In the absence of statistical similarity in aligned encoders like CLIP, we show that a possible matching of unaligned encoders exists without any training. We frame this as a seeded graph-matching problem exploiting the semantic similarity between graphs and propose two methods - a Fast Quadratic Assignment Problem optimization, and a novel localized CKA metric-based matching/retrieval. We demonstrate the effectiveness of this on several downstream tasks including cross-lingual, cross-domain caption matching and image classification.

Visual program synthesis is a promising approach to exploit the reasoning abilities of large language models for compositional computer vision tasks. Previous work has used few-shot prompting with frozen LLMs to synthesize visual programs.Training an LLM to write better visual programs is an attractive prospect, but it is unclear how to accomplish this.No dataset of visual programs for training exists, and acquisition of a visual program dataset cannot be easily crowdsourced due to the need for expert annotators.To get around the lack of direct supervision, we explore improving the program synthesis abilities of a LLM using feedback from interactive experience.We propose a method in which we exploit existing annotations for a vision-language task to improvise a coarse reward signal for that task, treat the LLM as a policy, and apply reinforced self-training to improve the visual program synthesis ability of the LLM for that task. We describe a series of experiments on object detection, compositional visual question answering, and image-text retrieval, and show that in each case, the self-trained LLM outperforms or performs on par with few-shot frozen LLMs that are an order of magnitude larger.

We study the visual semantic embedding problem for image-text matching. Most existing work utilizes a tailored cross-attention mechanism to perform local alignment across the two image and text modalities. This is computationally expensive, even though it is more powerful than the unimodal dual-encoder approach. This work introduces a dual-encoder image-text matching model, leveraging a scene graph to represent captions with nodes for objects and attributes interconnected by relational edges. Utilizing a graph attention network, our model efficiently encodes object-attribute and object-object semantic relations, resulting in a robust and fast-performing system. Representing caption as a scene graph offers the ability to utilize the strong relational inductive bias of graph neural networks to learn object-attribute and object-object relations effectively. To train the model, we propose losses that align the image and caption both at the holistic level (image-caption) and the local level (image-object entity), which we show is key to the success of the model. Our model is termed Composition model for Object Relations and Attributes, CORA. Experimental results on two prominent image-text retrieval benchmarks, Flickr30K and MS-COCO, demonstrate that CORA outperforms existing state-of-the-art computationally expensive cross-attention methods regarding recall score while achieving fast computation speed of the dual encoder.

Zero-shot referring expression comprehension aims at localizing bounding boxes in an image corresponding to provided textual prompts, which requires: (i) a fine-grained disentanglement of complex visual scene and textual context, and (ii) a capacity to understand relationships among disentangled entities. Unfortunately, existing large vision-language alignment (VLA) models, e.g., CLIP, struggle with both aspects so cannot be directly used for this task. To mitigate this gap, we leverage large foundation models to disentangle both images and texts into triplets in the format of (subject, predicate, object). After that, grounding is accomplished by calculating the structural similarity matrix between visual and textual triplets with a VLA model, and subsequently propagate it to an instance-level similarity matrix. Furthermore, to equip VLA models with the ability of relationship understanding, we design a triplet-matching objective to fine-tune the VLA models on a collection of curated dataset containing abundant entity relationships. Experiments demonstrate that our visual grounding performance increase of up to 19.5% over the SOTA zero-shot model on RefCOCO/+/g. On the more challenging Who’s Waldo dataset, our zero-shot approach achieves comparable accuracy to the fully supervised model

We introduce “HallusionBench,” a comprehensive benchmark designed for the evaluation of image-context reasoning. This benchmark presents significant challenges to advanced large visual-language models (LVLMs), such as GPT-4V(ision), Gemini Pro Vision, Claude 3, and LLaVA-1.5, by emphasizing nuanced understanding and interpretation of visual data. The benchmark comprises 346 images paired with 1129 questions, all meticulously crafted by human experts. We introduce a novel structure for these visual questions designed to establish control groups. This structure enables us to conduct a quantitative analysis of the models' response tendencies, logical consistency, and various failure modes. In our evaluation on HallusionBench, we benchmarked 15 different models, highlighting a 31.42% question-pair accuracy achieved by the state-of-the-art GPT-4V. Notably, all other evaluated models achieve accuracy below 16%. Moreover, our analysis not only highlights the observed failure modes, including language hallucination and visual illusion but also deepens an under standing of these pitfalls. Our comprehensive case studies within HallusionBench shed light on the challenges of hallucination and illusion in LVLMs. Based on these insights, we suggest potential pathways for their future improvement. The benchmark and codebase can be accessed at https://github.com/tianyilab/HallusionBench.

Understanding long, real-world videos requires modeling of long-range visual dependencies. To this end, we explore video-first architectures, building on the common paradigm of transferring large-scale, image--text models to video via shallow temporal fusion. However, we expose two limitations to the approach: (1) decreased spatial capabilities, likely due to poor video--language alignment in standard video datasets and (2) higher memory consumption, bottlenecking the number of frames that can be processed. To mitigate the memory bottleneck, we systematically analyze the memory/accuracy trade-off of various efficient methods: factorized attention, parameter-efficient image-to-video adaptation, input masking, and multi-resolution patchification. Surprisingly, simply masking large portions of the video (up to 75%) during contrastive pre-training proves to be one of the most robust ways to scale encoders to videos up to 4.3 minutes at 1 FPS. Our simple approach for training long video-to-text models, which scales to 1B parameters, does not add new architectural complexity and is able to outperform the popular paradigm of using much larger LLMs as an information aggregator over segment-based information on benchmarks with long-range temporal dependencies (YouCook2, EgoSchema).

Humans can easily solve multimodal tasks in context, with only a few demonstrations or simple instructions, which current multimodal systems largely struggle to imitate. In this work, we demonstrate that by effectively scaling up generative multimodal models, their task-agnostic in-context learning capabilities can be significantly enhanced.We introduce Emu2, a generative multimodal model with 37 billion parameters, which serves as a base model and general-purpose interface for a variety of multimodal tasks. Emu2 not only achieves strong performance in few-shot setting, but can also be instruct-tuned to follow specific instructions such as visual question answering and object-grounded image generation.Emu2 even emerges to solve tasks that require on-the-fly reasoning, such as visual prompting, which existing models are unlikely to handle. We identify additional tasks where Emu2's in-context learning can further improve, and discuss its broader societal impact.Our code and models will be made publicly available to facilitate future research.

What does learning to model relationships between strings teach Large Language Models (LLMs) about the visual world? We systematically evaluate LLMs’ abilities to generate and recognize an assortment of visual concepts of increasing complexity and then demonstrate how a preliminary visual representation learning system can be trained using models of text. As language models lack the ability to consume or output visual information as pixels, we use code to represent images in our study. Although LLM-generated images do not look like natural images, results on image generation and the ability of models to correct these generated images indicate that precise modeling of strings can teach language models about numerous aspects of the visual world. Furthermore, experiments on self-supervised visual representation learning, utilizing images generated with text models, highlight the potential to train vision models capable of making semantic assessments of natural images using just LLMs.

The combination of strong visual backbones and Large Language Model (LLM) reasoning has led to Large Multimodal Models (LMMs) becoming the current standard for a wide range of vision and language (VL) tasks. However, recent research has shown that even the most advanced LMMs still struggle to capture aspects of compositional visual reasoning, such as attributes and relationships between objects. One solution is to utilize scene graphs (SGs)---a formalization of objects and their relations and attributes that has been extensively used as a bridge between the visual and textual domains. Yet, scene graph data requires scene graph annotations, which are expensive to collect and thus not easily scalable. Moreover, finetuning an LMM based on SG data can lead to catastrophic forgetting of the pretraining objective. To overcome this, inspired by chain-of-thought methods, we propose Compositional Chain-of-Thought (CCoT), a novel zero-shot Chain-of-Thought prompting method that utilizes SG representations in order to extract compositional knowledge from an LMM. Specifically, we first generate an SG using the LMM, and then use that SG in the prompt to produce a response. Through extensive experiments, we find that the proposed CCoT approach not only improves LMM performance on several vision and language VL compositional benchmarks but also improves the performance of several popular LMMs on general multimodal benchmarks, without the need for fine-tuning or annotated ground-truth SGs. Code: https://github.com/chancharikmitra/CCoT

We explore the boundaries of scaling up a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. Our model advances the state-of-the-art on most vision-and-language benchmarks considered (20+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.

Automatic web navigation aims to build a web agent that can follow language instructions to execute complex and diverse tasks on real-world websites. Existing work primarily takes HTML documents as input, which define the contents and action spaces (i.e., actionable elements and operations) of webpages. Nevertheless, HTML documents may not provide a clear task-related context for each element, making it hard to select the right (sequence of) actions. In this paper, we propose to contextualize HTML elements through their "dual views" in webpage screenshots: each HTML element has its corresponding bounding box and visual content in the screenshot. We build upon the insight---web developers tend to arrange task-related elements nearby on webpages to enhance user experiences---and propose to contextualize each element with its neighbor elements, using both textual and visual features. The resulting representations of HTML elements are more informative for the agent to take actions. We validate our method on the recently released Mind2Web dataset, which features diverse navigation domains and tasks on real-world websites. Our method consistently outperforms the baseline in all the scenarios, including cross-task, cross-website, and cross-domain ones.

Understanding and reasoning about spatial relationships is crucial for Visual Question Answering (VQA) and robotics. Vision Language Models (VLMs) have shown impressive performance in some VQA benchmarks but struggle with 3D spatial reasoning, such as recognizing distances or size differences between physical objects. This limitation may stem from a lack of 3D spatial knowledge in their training data. To address this, we propose training VLMs with extensive spatial reasoning data from the internet. Our approach includes developing an automatic 3D spatial VQA data generation framework, capable of creating 2 billion VQA examples from 10 million real-world images. We explore various factors in the training process, such as data quality, training pipeline, and VLM architecture. Our work introduces the first Internet-scale 3D spatial reasoning dataset in metric space. By co-training a VLM with this dataset, we significantly improve its performance in both qualitative and quantitative spatial VQA. Additionally, this enhanced VLM enables new applications in chain-of-thought spatial reasoning and robotics, particularly in quantitative estimation.

Learning from seen attribute-object pairs to generalize to unseen compositions has been studied extensively in Compositional Zero-Shot Learning (CZSL). However, CZSL setup is still limited to seen attributes and objects, and cannot generalize to unseen concepts and their compositions. To overcome this limitation, we propose a new task, Open Vocabulary-Compositional Zero-shot Learning (OV-CZSL), where unseen attributes, objects, and unseen compositions are evaluated. To show that OV-CZSL is a challenging yet solvable problem, we propose three new benchmarks based on existing datasets MIT-States, C-GQA and VAW-CZSL along with new baselines and evaluation setup. We use language embeddings and external vocabulary with our novel neighborhood expansion loss to allow any method to learn semantic correlations between seen and unseen primitives.