Recent progress in text-driven 3D object stylization has been considerably promoted by the Contrastive Language-Image Pre-training (CLIP) model. However, the stylization of multi-object 3D scenes is impeded in that the image-text pairs used for pre-training CLIP mostly consist of an object. Meanwhile, the local details of multiple objects may be susceptible to omission due to the existing supervision manner primarily relying on coarse-grained contrast of image-text pairs. To overcome these challenges, we present a novel framework, dubbed TeMo, to parse multi-object 3D scenes and edit their styles under the contrast supervision at multiple levels. We first propose a Graph-based Cross Attention (GCA) module to distinguishably reinforce the features of 3D surface points. Particularly, a cross-modal graph is constructed to align the object points accurately and noun phrases decoupled from the 3D mesh and textual description. Then, we develop a Cross-Grained Contrast (CGC) supervision system, where a fine-grained loss between the words in the textual description and the randomly rendered images are constructed to complement the coarse-grained loss. Extensive experiments show that our method can synthesize desired styles and outperform the existing methods over a wide range of multi-object 3D meshes. Our codes and results will be made publicly available.

Event sensors offer high temporal resolution visual sensing, which makes them ideal for perceiving fast visual phenomena without suffering from motion blur. Certain applications in robotics and vision-based navigation require 3D perception of an object undergoing circular or spinning motion in front of a static camera, such as recovering the angular velocity and shape of the object. The setting is equivalent to observing a static object with an orbiting camera. In this paper, we propose event-based structure-from-orbit (eSfO), where the aim is to simultaneously reconstruct the 3D structure of a fast spinning object observed from a static event camera, and recover the equivalent orbital motion of the camera. Our contributions are threefold: since state-of-the-art event feature trackers cannot handle periodic self-occlusion due to the spinning motion, we develop a novel event feature tracker based on spatio-temporal clustering and data association that can better track the helical trajectories of valid features in the event data. The feature tracks are then fed to our novel factor graph-based structure-from-orbit back-end that calculates the orbital motion parameters (e.g., spin rate, relative rotational axis) that minimize the reprojection error. For evaluation, we produce a new event dataset of objects under spinning motion. Comparisons against ground truth indicate the efficacy of eSfO.

The rapid advancement of deep learning models often attributes to their ability to leverage massive training data. In contrast, such privilege has not yet fully benefited 3D deep learning, mainly due to the limited availability of large-scale 3D datasets. Merging multiple available data sources and letting them collaboratively train a single model is a potential solution. However, due to the large domain gap between 3D point cloud datasets, such mixed supervision could adversely affect the model's performance and lead to degenerated performance (i.e., negative transfer) compared to single-dataset training. In view of this challenge, we introduce Point Prompt Training (PPT), a novel framework for multi-dataset synergistic learning in the context of 3D representation learning that supports multiple pre-training paradigms. Based on this framework, we propose Prompt-driven Normalization, which adapts the model to different datasets with domain-specific prompts and Language-guided Categorical Alignment that decently unifies the multiple-dataset label spaces by leveraging the relationship between label text. Extensive experiments verify that PPT can overcome the negative transfer associated with synergistic learning and produce generalizable representations. Notably, it achieves state-of-the-art performance on each dataset using a single weight-shared model with supervised multi-dataset training. Moreover, when served as a pre-training framework, it outperforms other pre-training approaches regarding representation quality and attains remarkable state-of-the-art performance across over ten diverse downstream tasks spanning both indoor and outdoor 3D scenarios.

Photorealistic simulation plays a crucial role in applications such as autonomous driving, where advances in neural radiance fields (NeRFs) may allow better scalability through the automatic creation of digital 3D assets. However, reconstruction quality suffers on street scenes due to largely collinear camera motions and sparser samplings at higher speeds. On the other hand, the application often demands rendering from camera views that deviate from the inputs to accurately simulate behaviors like lane changes. In this paper, we propose several insights that allow a better utilization of Lidar data to improve NeRF quality on street scenes. First, our framework learns a geometric scene representation from Lidar, which are fused with the implicit grid-based representation for radiance decoding, thereby supplying stronger geometric information offered by explicit point cloud. Second, we put forth a robust occlusion-aware depth supervision scheme, which allows utilizing densified Lidar points by accumulation. Third, we generate augmented training views from Lidar points for further improvement. Our insights translate to largely improved novel view synthesis under real driving scenes.

The perception of 3D motion of surrounding traffic participants is crucial for driving safety. While existing works primarily focus on general large motions, we contend that the instantaneous detection and quantification of subtle motions is equally important as they indicate the nuances in driving behavior that may be safety critical, such as behaviors near a stop sign of parking positions. We delve into this under-explored task, examining its unique challenges and developing our solution, accompanied by a carefully designed benchmark. Specifically, due to the lack of correspondences between consecutive frames of sparse Lidar point clouds, static objects might appear to be moving – the so-called swimming effect. This intertwines with the true object motion, thereby posing ambiguity in accurate estimation, especially for subtle motion. To address this, we propose to leverage local occupancy completion of object point clouds to densify the shape cue, and mitigate the impact of swimming artifacts. The occupancy completion is learned in an end-to-end fashion together with the detection of moving objects and the estimation of their motion, instantaneously as soon as objects start to move. Extensive experiments demonstrate superior performance compared to standard 3D motion estimation approaches, particularly highlighting our method's specialized treatment of subtle motion.

Implicit neural SLAM has achieved remarkable progress recently. Nevertheless, existing methods face significant challenges in non-ideal scenarios, such as motion blur or lighting variation, which often leads to issues like convergence failures, localization drifts, and distorted mapping. To address these challenges, we propose EN-SLAM, the first event-RGBD implicit neural SLAM framework, which effectively leverages the high rate and high dynamic range advantages of event data for tracking and mapping. Specifically, EN-SLAM proposes a differentiable CRF (Camera Response Function) rendering technique to generate distinct RGB and event camera data via a shared radiance field, which is optimized by learning a unified implicit representation with the captured event and RGBD supervision. Moreover, based on the temporal difference property of events, we propose a temporal aggregating optimization strategy for the event joint tracking and global bundle adjustment, capitalizing on the consecutive difference constraints of events, significantly enhancing tracking accuracy and robustness. Finally, we construct the simulated dataset DEV-Indoors and real captured dataset DEV-Reals containing 6 scenes, 17 sequences with practical motion blur and lighting changes for evaluations. Experimental results show that our method outperforms the SOTA methods in both tracking ATE and mapping ACC with a real-time 17 FPS in various challenging environments. Project page: https://delinqu.github.io/EN-SLAM.

In this paper, we introduce \textbf{GS-SLAM} that first utilizes 3D Gaussian representation in the Simultaneous Localization and Mapping (SLAM) system. It facilitates a better balance between efficiency and accuracy. Compared to recent SLAM methods employing neural implicit representations, our method utilizes a real-time differentiable splatting rendering pipeline that offers significant speedup to map optimization and RGB-D rendering. Specifically, we propose an adaptive expansion strategy that adds new or deletes noisy 3D Gaussians in order to efficiently reconstruct new observed scene geometry and improve the mapping of previously observed areas. This strategy is essential to extend 3D Gaussian representation to reconstruct the whole scene rather than synthesize a static object in existing methods. Moreover, in the pose tracking process, an effective coarse-to-fine technique is designed to select reliable 3D Gaussian representations to optimize camera pose, resulting in runtime reduction and robust estimation. Our method achieves competitive performance compared with existing state-of-the-art real-time methods on the Replica, TUM-RGBD datasets. Project page: \href{https://gs-slam.github.io/}{https://gs-slam.github.io/}.

Multi-instance point cloud registration estimates the poses of multiple source point cloud instances in a target point cloud. Solving this problem relies on first extracting point correspondences. However, existing methods treat the target point cloud as a whole, neglecting the independence of instances. As a result, point features could be easily poluted by those from background or other instances, leading to inaccurate correspondences and missing instances, especially in cluttered scenes. In this work, we propose Multi-Instance REgistration TRansformer, a coarse-to-fine framework to directly extract correspondences and estimate the transformation for each instance. In the coarse level, our method jointly learns instance-aware superpoint features and predicts local instance masks. Benefiting from the instance masks, the influence from outside of instance is alleviated, such that highly reliable superpoint correspondences are extracted. The superpoint correspondences are further extended to instance candidates in the fine level according to the instance masks. At last, an efficient candidate selection and refinement algorithm is devised to obtain the final registrations. Extensive experiments on two benchmarks have demonstrated the efficacy of our design. MIRETR outperforms the previous state-of-the-art by over 30.22 points on F1 score on the challenging ROBI benchmark. Our code and models will be released.

We introduce MeshGPT, a new approach for generating triangle meshes that reflects the compactness typical of artist-created meshes, in contrast to dense triangle meshes extracted by iso-surfacing methods from neural fields. Inspired by recent advances in powerful large language models, we adopt a sequence-based approach to autoregressively generate triangle meshes as sequences of triangles. We first learn a vocabulary of latent quantized embeddings, using graph convolutions, which inform these embeddings of the local mesh geometry and topology. These embeddings are sequenced and decoded into triangles by a decoder, ensuring that they can effectively reconstruct the mesh. A transformer is then trained on this learned vocabulary to predict the index of the next embedding given previous embeddings. Once trained, our model can be autoregressively sampled to generate new triangle meshes, directly generating compact meshes with sharp edges, more closely imitating the efficient triangulation patterns of human-crafted meshes. MeshGPT demonstrates a notable improvement over state of the art mesh generation methods, with a 9% increase in shape coverage and a 30-point enhancement in FID scores across various categories.

We introduce a new system for Multi-Session SLAM, which tracks camera motion across multiple disjoint videos under a single global reference. Our approach couples the prediction of optical flow with solver layers to estimate camera pose. It is trained end-to-end using a novel differentiable solver for wide-baseline two-view pose. The full system can connect disjoint sequences, perform visual odometry, and efficient global-optimization/loop-closure. Compared to other Multi-Session SLAM approaches, our design is more accurate and robust to catastrophic failures.

We present SHINOBI, an end-to-end framework for reconstruction of shape, material and illumination from images captured with varying lighting, pose and background. Inverse rendering of an object based on unconstrained image collections is a long-standing challenge in computer vision and graphics and requires a joint optimization over shape, radiance, and pose. We show that an implicit shape representation based on a multi-resolution hash encoding enables fast and robust shape reconstruction with joint camera alignment optimization that outperforms prior work. Further, to enable the editing of illumination and object reflectance (i.e. material) we jointly optimize BRDF and illumination together with the object's shape. Our method is class-agnostic and works on in-the-wild image collections of objects to produce relightable 3D assets for several use-cases such as AR/VR.

Neural radiance fields provide state-of-the-art view synthesis quality but tend to be slow to render. One reason is that they make use of volume rendering, thus requiring many samples (and model queries) per ray at render time. Although this representation is flexible and easy to optimize, most real-world objects can be modeled more efficiently with surfaces instead of volumes, requiring far fewer samples per ray. This observation has spurred considerable progress in surface representations such as signed distance functions, but these may struggle to model semi-opaque and thin structures. We propose a method, HybridNeRF, that leverages the strengths of both representations by rendering most objects as surfaces while modeling the (typically) small fraction of challenging regions volumetrically.We evaluate HybridNeRF against the challenging Eyeful Tower dataset along with other commonly used view synthesis datasets. When comparing to state-of-the-art baselines, including recent rasterization-based approaches, we improve error rates by 15-30% while achieving real-time framerates (at least 36 FPS) for virtual-reality resolutions (2K x 2K).

Neural implicit scene representations have recently shown encouraging results in dense visual SLAM. However, existing methods produce low-quality scene reconstruction and low-accuracy localization performance when scaling up to large indoor scenes and long sequences. These limitations are mainly due to their single, global radiance field with finite capacity, which does not adapt to large scenarios. Their end-to-end pose networks are also not robust enough with the growth of cumulative errors in large scenes. To this end, we present PLGSLAM, a neural visual SLAM system which performs high-fidelity surface reconstruction and robust camera tracking in real time. To handle large-scale indoor scenes, PLGSLAM proposes a progressive scene representation method which dynamically allocates new local scene representation trained with frames within a local sliding window. This allows us to scale up to larger indoor scenes and improves robustness (even under pose drifts). In local scene representation, PLGSLAM utilizes tri-planes for local high-frequency features. We also incorporate multi-layer perceptron (MLP) networks for the low-frequency feature, smoothness, and scene completion in unobserved areas. Moreover, we propose local-to-global bundle adjustment method with a global keyframe database to address the increased pose drifts on long sequences. Experimental results demonstrate that PLGSLAM achieves state-of-the-art scene reconstruction results and tracking performance across various datasets and scenarios (both in small and large-scale indoor environments). The code will be open-sourced upon paper acceptance.

Extensions of Neural Radiance Fields (NeRFs) to model dynamic scenes have enabled their near photo-realistic, free-viewpoint rendering. Although these methods have shown some potential in creating immersive experiences, two drawbacks limit their ubiquity: (i) a significant reduction in reconstruction quality when the computing budget is limited, and (ii) a lack of semantic understanding of the underlying scenes. To address these issues, we introduce Gear-NeRF, which leverages semantic information from powerful image segmentation models. Our approach presents a principled way for learning a spatio-temporal (4D) semantic embedding, based on which we introduce the concept of gears to allow for stratified modeling of dynamic regions of the scene based on the extent of their motion. Such differentiation allows us to adjust the spatio-temporal sampling resolution for each region in proportion to its motion scale, achieving more photo-realistic dynamic novel view synthesis. At the same time, almost for free, our approach enables free-viewpoint tracking of objects of interest -- a functionality not yet achieved by existing NeRF-based methods. Empirical studies validate the effectiveness of our method, where we achieve state-of-the-art rendering and tracking performance on multiple challenging datasets. The project page is available at: https://merl.com/research/highlights/gear-nerf.

We present a new approach, termed GPS-Gaussian, for synthesizing novel views of a character in a real-time manner. The proposed method enables 2K-resolution rendering under a sparse-view camera setting. Unlike the original Gaussian Splatting or neural implicit rendering methods that necessitate per-subject optimizations, we introduce Gaussian parameter maps defined on the source views and regress directly Gaussian Splatting properties for instant novel view synthesis without any fine-tuning or optimization. To this end, we train our Gaussian parameter regression module on a large amount of human scan data, jointly with a depth estimation module to lift 2D parameter maps to 3D space. The proposed framework is fully differentiable and experiments on several datasets demonstrate that our method outperforms state-of-the-art methods while achieving an exceeding rendering speed.

3D shape generation from text is a fundamental task in 3D representation learning. The text-shape pairs exhibit a hierarchical structure, where a general text like "chair" covers all 3D shapes of the chair, while more detailed prompts refer to more specific shapes. Furthermore, both text and 3D shapes are inherently hierarchical structures. However, existing Text2Shape methods, such as SDFusion, do not exploit that. In this work, we propose HyperSDFusion, a dual-branch diffusion model that generates 3D shapes from a given text. Since hyperbolic space is suitable for handling hierarchical data, we propose to learn the hierarchical representations of text and 3D shapes in hyperbolic space. First, we introduce a hyperbolic text-image encoder to learn the sequential and multi-modal hierarchical features of text in hyperbolic space. In addition, we design a hyperbolic text-graph convolution module to learn the hierarchical features of text in hyperbolic space. In order to fully utilize these text features, we introduce a dual-branch structure to embed text features in 3D feature space. At last, to endow the generated 3D shapes with a hierarchical structure, we devise a hyperbolic hierarchical loss. Our method is the first to explore the hyperbolic hierarchical representation for text-to-shape generation. Experimental results on the existing text-to-shape paired dataset, Text2Shape, achieved state-of-the-art results.

Stereo matching methods based on iterative optimization, like RAFT-Stereo and IGEV-Stereo, have evolved into a cornerstone in the field of stereo matching. However, these methods struggle to simultaneously capture high-frequency information in edges and low-frequency information in smooth regions due to the fixed receptive field. As a result, they tend to lose details, blur edges, and produce false matches in textureless areas. In this paper, we propose Selective Recurrent Unit (SRU), a novel iterative update operator for stereo matching. The SRU module can adaptively fuse hidden disparity information at multiple frequencies for edge and smooth regions. To perform adaptive fusion, we introduce a new Contextual Spatial Attention (CSA) module to generate attention maps as fusion weights. The SRU empowers the network to aggregate hidden disparity information across multiple frequencies, mitigating the risk of vital hidden disparity information loss during iterative processes. To verify SRU's universality, we apply it to representative iterative stereo matching methods, collectively referred to as Selective-Stereo. Our Selective-Stereo ranks $1^{st}$ on KITTI 2012, KITTI 2015, ETH3D, and Middlebury leaderboards among all published methods. The source code will be available upon the publicity of the paper.

Modeling animatable human avatars from RGB videos is a long-standing and challenging problem. Recent works usually adopt MLP-based neural radiance fields (NeRF) to represent 3D humans, but it remains difficult for pure MLPs to regress pose-dependent garment details. To this end, we introduce Animatable Gaussians, a new avatar representation that leverages powerful 2D CNNs and 3D Gaussian splatting to create high-fidelity avatars. To associate 3D Gaussians with the animatable avatar, we learn a parametric template from the input videos, and then parameterize the template on two front \& back canonical Gaussian maps where each pixel represents a 3D Gaussian. The learned template is adaptive to the wearing garments for modeling looser clothes like dresses. Such template-guided 2D parameterization enables us to employ a powerful StyleGAN-based CNN to learn the pose-dependent Gaussian maps for modeling detailed dynamic appearances. Furthermore, we introduce a pose projection strategy for better generalization given novel poses. Overall, our method can create lifelike avatars with dynamic, realistic and generalized appearances. Experiments show that our method outperforms other state-of-the-art approaches. Code will be public.

A recent trend among generalizable novel view synthesis methods is to learn a rendering operator acting over single camera rays. This approach is promising because it removes the need for explicit volumetric rendering, but it effectively treats target images as collections of independent pixels. Here, we propose to learn a global rendering operator acting over all camera rays jointly. We show that the right representation to enable such rendering is a 5-dimensional plane sweep volume consisting of the projection of the input images on a set of planes facing the target camera. Based on this understanding, we introduce our Convolutional Global Latent Renderer (ConvGLR), an efficient convolutional architecture that performs the rendering operation globally in a low-resolution latent space. Experiments on various datasets under sparse and generalizable setups show that our approach consistently outperforms existing methods by significant margins.

We have recently seen tremendous progress in photo-real human modeling and rendering. Yet, efficiently rendering realistic human performance and integrating it into the rasterization pipeline remains challenging. In this paper, we present HiFi4G, an explicit and compact Gaussian-based approach for high-fidelity human performance rendering from dense footage. Our core intuition is to marry the 3D Gaussian representation with non-rigid tracking, achieving a compact and compression-friendly representation. We first propose a dual-graph mechanism to obtain motion priors, with a coarse deformation graph for effective initialization and a fine-grained Gaussian graph to enforce subsequent constraints. Then, we utilize a 4D Gaussian optimization scheme with adaptive spatial-temporal regularizers to effectively balance the non-rigid prior and Gaussian updating. We also present a companion compression scheme with residual compensation for immersive experiences on various platforms. It achieves a substantial compression rate of approximately 25 times, with less than 2MB of storage per frame. Extensive experiments demonstrate the effectiveness of our approach, which significantly outperforms existing approaches in terms of optimization speed, rendering quality, and storage overhead.

Estimating disparities in challenging areas is difficult and limits the performance of stereo matching models. In this paper, we exploit local structure information (LSI) to enhance stereo matching. Specifically, our LSI comprises a series of key elements, including the slant plane (parameterised by disparity gradients), disparity offset details and neighbouring relations. This LSI empowers our method to effectively handle intricate structures, including object boundaries and curved surfaces. We bootstrap the LSI from monocular depth and subsequently iteratively refine it to better capture the underlying scene geometry constraints. Building upon the LSI, we introduce the Local Structure-Guided Propagation (LSGP), which enhances the disparity initialization, optimization, and refinement processes. By combining LSGP with a Gated Recurrent Unit (GRU), we present our novel stereo matching method, referred to as $\textbf{Lo}$cal $\textbf{S}$tructure-guided stereo matching (LoS). Remarkably, LoS achieves top-ranking results on four widely recognized public benchmark datasets (ETH3D, Middlebury, KITTI 15 & 12), demonstrating the superior capabilities of our proposed model.

In the realm of computer vision and robotics, embodied agents are expected to explore their environment and carry out human instructions. This necessitates the ability to fully understand 3D scenes given their first-person observations and contextualize them into language for interaction. However, traditional research focuses more on scene-level input and output setups from a global view. To address the gap, we introduce EmbodiedScan, a multi-modal, ego-centric 3D perception dataset and benchmark for holistic 3D scene understanding. It encompasses over 5k scans encapsulating 1M ego-centric RGB-D views, 1M language prompts, 160k 3D-oriented boxes spanning over 760 categories, some of which partially align with LVIS, and dense semantic occupancy with 80 common categories. Building upon this database, we introduce a baseline framework named Embodied Perceptron. It is capable of processing an arbitrary number of multi-modal inputs and demonstrates remarkable 3D perception capabilities, both within the two series of benchmarks we set up, i.e., fundamental 3D perception tasks and language-grounded tasks, and in the open world.

The main function of depth completion is to compensate for an insufficient and unpredictable number of sparse depth measurements of hardware sensors. However, existing research on depth completion assumes that the sparsity --- the number of points or LiDAR lines --- is fixed for training and testing. Hence, the completion performance drops severely when the number of sparse depths changes significantly. To address this issue, we propose the sparsity-adaptive depth refinement (SDR) framework, which refines monocular depth estimates using sparse depth points. For SDR, we propose the masked spatial propagation network (MSPN) to perform SDR with a varying number of sparse depths effectively by gradually propagating sparse depth information throughout the entire depth map. Experimental results demonstrate that MPSN achieves state-of-the-art performance on both SDR and conventional depth completion scenarios.

Deep neural networks have demonstrated remarkable performance in point cloud classification. However, previous works show they are vulnerable to adversarial perturbations that can manipulate their predictions. Given the distinctive modality of point clouds, various attack strategies have emerged, posing challenges for existing defenses to achieve effective generalization. In this study, we for the first time introduce causal modeling to enhance the robustness of point cloud classification models. Our insight is from the observation that adversarial examples closely resemble benign point clouds from the human perspective. In our causal modeling, we incorporate two critical variables, the structural information, (standing for the key feature leading to the classification) and the hidden confounders, (standing for the noise interfering with the classification). The resulting overall framework CausalPC consists of three sub-modules to identify the causal effect for robust classification. The framework is model-agnostic and adaptable for integration with various point cloud classifiers. Our approach significantly improves the adversarial robustness of three mainstream point cloud classification models on two benchmark datasets. For instance, the classification accuracy for DGCNN on ModelNet40 increases from 29.2% to 72.0% with CausalPC, whereas the best-performing baseline achieves only 42.4%.

Feature matching is an important computer vision task that involves estimating correspondences between two images of a 3D scene, and dense methods estimate all such correspondences. The aim is to learn a robust model, i.e., a model able to match under challenging real-world changes. In this work, we propose such a model, leveraging frozen pretrained features from the foundation model DINOv2. Although these features are significantly more robust than local features trained from scratch, they are inherently coarse. We therefore combine them with specialized ConvNet fine features, creating a precisely localizable feature pyramid. To further improve robustness, we propose a tailored transformer match decoder that predicts anchor probabilities, which enables it to express multimodality. Finally, we propose an improved loss formulation through regression-by-classification with subsequent robust regression. We conduct a comprehensive set of experiments that show that our method, RoMa, achieves significant gains, setting a new state-of-the-art. In particular, we achieve a 36% improvement on the extremely challenging WxBS benchmark. Code is provided at github.com/Parskatt/RoMa.

In this era, the success of large language models and text-to-image models can be attributed to the driving force of large-scale datasets. However, in the realm of 3D vision, while remarkable progress has been made with models trained on large-scale synthetic and real-captured object data like Objaverse and MVImgNet, a similar level of progress has not been observed in the domain of human-centric tasks partially due to the lack of a large-scale human dataset. Existing datasets of high-fidelity 3D human capture continue to be mid-sized due to the significant challenges in acquiring large-scale high-quality 3D human data. To bridge this gap, we present MVHumanNet, a dataset that comprises multi-view human action sequences of 4,500 human identities. The primary focus of our work is on collecting human data that features a large number of diverse identities and everyday clothing using a multi-view human capture system, which facilitates easily scalable data collection. Our dataset contains 9,000 daily outfits, 60,000 motion sequences and 645 million frames with extensive annotations, including human masks, camera parameters, 2D and 3D keypoints, SMPL/SMPLX parameters, and corresponding textual descriptions. To explore the potential of MVHumanNet in various 2D and 3D visual tasks, we conducted pilot studies on view-consistent action recognition, human NeRF reconstruction, text-driven view-unconstrained human image generation, as well as 2D view-unconstrained human image and 3D avatar generation. Extensive experiments demonstrate the performance improvements and effective applications enabled by the scale provided by MVHumanNet. As the current largest-scale 3D human dataset, we hope that the release of MVHumanNet data with annotations will foster further innovations in the domain of 3D human-centric tasks at scale.

Advancements in 3D Gaussian Splatting have significantly accelerated 3D reconstruction and generation. However, it may require a large number of Gaussians, which creates a substantial memory footprint. This paper introduces GES (Generalized Exponential Splatting), a novel representation that employs Generalized Exponential Function (GEF) to model 3D scenes, requiring far fewer particles to represent a scene and thus significantly outperforming Gaussian Splatting methods in efficiency with a plug-and-play replacement ability for Gaussian-based utilities. GES is validated theoretically and empirically in both principled 1D setup and realistic 3D scenes. It is shown to represent signals with sharp edges more accurately, which are typically challenging for Gaussians due to their inherent low-pass characteristics. Our empirical analysis demonstrates that GEF outperforms Gaussians in fitting natural-occurring signals (E.g. squares, triangles, parabolic signals), thereby reducing the need for extensive splitting operations that increase the memory footprint of Gaussian Splatting. With the aid of a frequency-modulated loss, GES achieves competitive performance in novel-view synthesis benchmarks while requiring less than half the memory storage of Gaussian Splatting and increasing the rendering speed by up to 39%. The code is available on the project website https://abdullahamdi.com/ges .

We propose a lightweight and scalable Regional Point-Language Contrastive learning framework, namely RegionPLC, for open-world 3D scene understanding, aiming to identify and recognize open-set objects and categories. Specifically, based on our empirical studies, we introduce a 3D-aware SFusion strategy that fuses 3D vision-language pairs derived from multiple 2D foundation models, yielding high-quality, dense region-level language descriptions without human 3D annotations. Subsequently, we devise a region-aware point-discriminative contrastive learning objective to enable robust and effective 3D learning from dense regional language supervision. We carry out extensive experiments on ScanNet, ScanNet200, and nuScenes datasets, and our model outperforms prior 3D open-world scene understanding approaches by an average of 17.2\% and 9.1\% for semantic and instance segmentation, respectively, while maintaining greater scalability and lower resource demands. Furthermore, our method has the flexibility to be effortlessly integrated with language models to enable open-ended grounded 3D reasoning without extra task-specific training. Code will be released.

We present NeLF-Pro, a novel representation to model and reconstruct light fields in diverse natural scenes that vary in extent and spatial granularity. In contrast to previous fast reconstruction methods that represent the 3D scene globally, we model the light field of a scene as a set of local light field feature probes, parameterized with position and multi-channel 2D feature maps. Our central idea is to bake the scene's light field into spatially varying learnable representations and to query point features by weighted blending of probes close to the camera - allowing for mipmap representation and rendering. We introduce a novel vector-matrix-matrix (VMM) factorization technique that effectively represents the light field feature probes as products of core factors (i.e., VM) shared among local feature probes, and a basis factor (i.e., M) - efficiently encoding internal relationships and patterns within the scene.Experimentally, we demonstrate that NeLF-Pro significantly boosts the performance of feature grid-based representations, and achieves fast reconstruction with better rendering quality while maintaining compact modeling. Project page: sinoyou.github.io/nelf-pro

Visual odometry estimates the motion of a moving camera based on visual input. Existing methods, mostly focusing on two-view point tracking, often ignore the rich temporal context in the image sequence, thereby overlooking the global motion patterns and providing no assessment of the full trajectory reliability. These shortcomings hinder performance in scenarios with occlusion, dynamic objects, and low-texture areas. To address these challenges, we present the Long-term Effective Any Point Tracking (LEAP) module. LEAP innovatively combines visual, inter-track, and temporal cues with mindfully selected anchors for dynamic track estimation. Moreover, LEAP's temporal probabilistic formulation integrates distribution updates into a learnable iterative refinement module to reason about point-wise uncertainty. Based on these traits, we develop LEAP-VO, a robust visual odometry system adept at handling occlusions and dynamic scenes. Our mindful integration showcases a novel practice by employing long-term point tracking as the front-end. Extensive experiments demonstrate that the proposed pipeline significantly outperforms existing baselines across various visual odometry benchmarks.

Estimating relative camera poses between images has been a central problem in computer vision. Methods that find correspondences and solve for the fundamental matrix offer high precision in most cases. Conversely, methods predicting pose directly using neural networks are more robust to limited overlap and can infer absolute translation scale, but at the expense of reduced precision. We show how to combine the best of both methods; our approach yields results that are both precise and robust, while also accurately inferring translation scales. At the heart of our model lies a Transformer that (1) learns to balance between solved and learned pose estimations, and (2) provides a prior to guide a solver. A comprehensive analysis supports our design choices and demonstrates that our method adapts flexibly to various feature extractors and correspondence estimators, showing state-of-the-art performance in 6DoF pose estimation on Matterport3D, InteriorNet, StreetLearn, and Map-Free Relocalization.

The image matching field has been witnessing a continuous emergence of novel learnable feature matching techniques, with ever-improving performance on conventional benchmarks. However, our investigation shows that despite these gains, their potential for real-world applications is restricted by their limited generalization capabilities to novel image domains. In this paper, we introduce OmniGlue, the first learnable image matcher that is designed with generalization as a core principle. OmniGlue leverages broad knowledge from a vision foundation model to guide the feature matching process, boosting generalization to domains not seen at training time. Additionally, we propose a novel keypoint position-guided attention mechanism which disentangles spatial and appearance information, leading to enhanced matching descriptors. We perform comprehensive experiments on a suite of $6$ datasets with varied image domains, including scene-level, object-centric and aerial images. OmniGlue's novel components lead to relative gains on unseen domains of 18.8% with respect to a directly comparable reference model, while also outperforming the recent LightGlue method by 10.1% relatively. Code and model will be released.

We introduce Gaussian Articulated Template Model (GART), an explicit, efficient, and expressive representation for non-rigid articulated subject capturing and rendering from monocular videos. GART utilizes a mixture of moving 3D Gaussians to explicitly approximate a deformable subject’s geometry and appearance. It takes advantage of a categorical template model prior (SMPL, SMAL, etc.) with learnable forward skinning while further generalizing to more complex non-rigid deformations with novel latent bones. GART can be reconstructed via differentiable rendering from monocular videos in seconds or minutes and rendered in novel poses faster than 150fps.

We propose CG-HOI, the first method to address the task of generating dynamic 3D human-object interactions (HOIs) from text. We model the motion of both human and object in an interdependent fashion, as semantically rich human motion rarely happens in isolation without any interactions. Our key insight is that explicitly modeling contact between the human body surface and object geometry can be used as strong proxy guidance, both during training and inference. Using this guidance to bridge human and object motion enables generating more realistic and physically plausible interaction sequences, where the human body and corresponding object move in a coherent manner. Our method first learns to model human motion, object motion, and contact in a joint diffusion process, inter-correlated through cross-attention. We then leverage this learned contact for guidance during inference to synthesize realistic and coherent HOIs. Extensive evaluation shows that our joint contact-based human-object interaction approach generates realistic and physically plausible sequences, and we show two applications highlighting the capabilities of our method. Conditioned on a given object trajectory, we can generate the corresponding human motion without re-training, demonstrating strong human-object interdependency learning. Our approach is also flexible, and can be applied to static real-world 3D scene scans.

We present a generative approach to forecast long-term future human behavior in 3D, requiring only weak supervision from readily available 2D human action data. This is a fundamental task enabling many downstream applications. The required ground-truth data is hard to capture in 3D (mocap suits, expensive setups) but easy to acquire in 2D (simple RGB cameras). Thus, we design our method to only require 2D RGB data at inference time while being able to generate 3D human motion sequences. We use a differentiable 2D projection scheme in an autoregressive manner for weak supervision, and an adversarial loss for 3D regularization.Our method predicts long and complex human behavior sequences (e.g., cooking, assembly) consisting of multiple sub-actions. We tackle this in a semantically hierarchical manner, jointly predicting high-level coarse action labels together with their low-level fine-grained realizations as characteristic 3D human poses. We observe that these two action representations are coupled in nature, and joint prediction benefits both action and pose forecasting. Our experiments demonstrate the complementary nature of joint action and 3D pose prediction: our joint approach outperforms each task treated individually, enables robust longer-term sequence prediction, and improves over alternative approaches to forecast actions and characteristic 3D poses.

Diffusion models trained on large-scale text-image datasets have demonstrated a strong capability of controllable high-quality image generation from arbitrary text prompts. However, the generation quality and generalization ability of 3D diffusion models is hindered by the scarcity of high-quality and large-scale 3D datasets. In this paper, we present PI3D, a framework that fully leverages the pre-trained text-to-image diffusion models' ability to generate high-quality 3D shapes from text prompts in minutes. The core idea is to connect the 2D and 3D domains by representing a 3D shape as a set of Pseudo RGB Images. We fine-tune an existing text-to-image diffusion model to produce such pseudo-images using a small number of text-3D pairs. Surprisingly, we find that it can already generate meaningful and consistent 3D shapes given complex text descriptions. We further take the generated shapes as the starting point for a lightweight iterative refinement using score distillation sampling to achieve high-quality generation under a low budget. PI3D generates a single 3D shape from text in only 3 minutes and the quality is validated to outperform existing 3D generative models by a large margin.

An effective pre-training framework with universal 3D representations is extremely desired in perceiving large-scale dynamic scenes. However, establishing such an ideal framework that is both task-generic and label-efficient poses a challenge in unifying the representation of the same primitive across diverse scenes. The current contrastive 3D pre-training methods typically follow a frame-level consistency, which focuses on the 2D-3D relationships in each detached image. Such inconsiderate consistency greatly hampers the promising path of reaching an universal pre-training framework: (1) The cross-scene semantic self-conflict, {\textit i.e.}, the intense collision between primitive segments of the same semantics from different scenes; (2) Lacking a globally unified bond that pushes the cross-scene semantic consistency into 3D representation learning. To address above challenges, we propose a CSC framework that puts a scene-level semantic consistency in the heart, bridging the connection of the similar semantic segments across various scenes. To achieve this goal, we combine the coherent semantic cues provided by the vision foundation model and the knowledge-rich cross-scene prototypes derived from the complementary multi-modality information. These allow us to train a universal 3D pre-training model that facilitates various downstream tasks with less fine-tuning efforts. Empirically, we achieve consistent improvements over SOTA pre-training approaches in semantic segmentation (+1.4\% mIoU), object detection (+1.0\% mAP), and panoptic segmentation (+3.0\% PQ) using their task-specific 3D network on nuScenes. We will release our code, hoping to inspire future research.

The autonomous driving community has shown significant interest in 3D occupancy prediction, driven by its exceptional geometric perception and general object recognition capabilities. To achieve this, current works try to construct a Tri-Perspective View (TPV) or Occupancy (OCC) representation extending from the Bird-Eye-View perception. However, compressed views like TPV representation lose 3D geometry information while raw and sparse OCC representation requires heavy but reducant computational costs. To address the above limitations, we propose Compact Occupancy TRansformer (COTR), with a geometry-aware occupancy encoder and a semantic-aware group decoder to reconstruct a compact 3D OCC representation. The occupancy encoder first generates a compact geometrical OCC feature through efficient explicit-implicit view transformation. Then, the occupancy decoder further enhances the semantic discriminability of the compact OCC representation by a coarse-to-fine semantic grouping strategy. Empirical experiments show that there are evident performance gains across multiple baselines, e.g., COTR outperforms baselines with a relative improvement of 8%-15%, demonstrating the superiority of our method.

3D occupancy prediction is an important task for the robustness of vision-centric autonomous driving, which aims to predict whether each point is occupied in the surrounding 3D space. Existing methods usually require 3D occupancy labels to produce reasonable results. However, it is very laborious to annotate the occupancy status of each voxel. In this paper, we propose SelfOcc to explore a self-supervised way to learn 3D occupancy using only video sequences. We first transform the images into the bird's eye view (BEV) or tri-perspective view (TPV) space to obtain 3D representation of the scene. We directly impose constraints on the 3D representations by treating them as a neural radiance field. We can then render 2D images of previous and future frames as self-supervision signals to learn the 3D representations. Our SelfOcc outperforms the previous best method SceneRF by 58.7% using a single frame as input on SemanticKITTI and is the first work that produces meaningful 3D occupancy for surround cameras on Occ3D. As a bonus, SelfOcc can also produce high-quality depth and achieves state-of-the-art results on novel depth synthesis, monocular depth estimation, and surround-view depth estimation on the SemanticKITTI, KITTI-2015, and nuScenes, respectively.

3D instance segmentation is fundamental to geometric understanding of the world around us. Existing methods for instance segmentation of 3D scenes rely on supervision from expensive, manual 3D annotations.We propose UnScene3D, the first fully unsupervised 3D learning approach for class-agnostic 3D instance segmentation of indoor scans. UnScene3D first generates pseudo masks by leveraging self-supervised color and geometry features to find potential object regions. We operate on a basis of 3D segment primitives, enabling efficient representation and learning on high-resolution 3D data. The coarse proposals are then refined through self-training our model on its predictions. Our approach improves over state-of-the-art unsupervised 3D instance segmentation methods by more than 300% Average Precision score, demonstrating effective instance segmentation even in challenging, cluttered 3D scenes.

This paper studies the problem of structured 3D reconstruction using wireframes that consist of line segments and junctions, focusing on the computation of structured boundary geometries of scenes. Instead of leveraging matching-based solutions from 2D wireframes (or line segments) for 3D wireframe reconstruction as done in prior arts, we present NEAT, a rendering-distilling formulation using neural fields to represent 3D line segments with 2D observations, and bipartite matching for perceiving and distilling of a sparse set of 3D global junctions. The proposed NEAT enjoys the joint optimization of the neural fields and the global junctions from scratch, using view-dependent 2D observations without precomputed cross-view feature matching. Comprehensive experiments on the DTU and BlendedMVS datasets demonstrate our NEAT's superiority over state-of-the-art alternatives for 3D wireframe reconstruction. Moreover, the distilled 3D global junctions by NEAT, are a better initialization than SfM points, for the recently-emerged 3D Gaussian Splatting for high-fidelity novel view synthesis using about 20 times fewer initial 3D points. Project page: https://xuenan.net/neat.

Neural Radiance Field (NeRF) has been widely recognized for its excellence in novel view synthesis and 3D scene reconstruction. However, their effectiveness is inherently tied to the assumption of static scenes, rendering them susceptible to undesirable artifacts when confronted with transient distractors such as moving objects or shadows. In this work, we propose a novel paradigm, namely ``Heuristics-Guided Segmentation'' (HuGS), which significantly enhances the separation of static scenes from transient distractors by harmoniously combining the strengths of hand-crafted heuristics and state-of-the-art segmentation models, thus significantly transcending the limitations of previous solutions. Furthermore, we delve into the meticulous design of heuristics, introducing a seamless fusion of Structure-from-Motion (SfM)-based heuristics and color residual heuristics, catering to a diverse range of texture profiles. Extensive experiments demonstrate the superiority and robustness of our method in mitigating transient distractors for NeRFs trained in non-static scenes. Project page: \url{https://cnhaox.github.io/NeRF-HuGS/}

We propose a novel method for 3D point cloud action recognition. Understanding human actions in RGB videos has been widely studied in recent years, however, its 3D point cloud counterpart remains under-explored. This is mostly due to the inherent limitation of the point cloud data modality---lack of structure, permutation invariance, and varying number of points---which makes it difficult to learn a spatio-temporal representation. To address this limitation, we propose the 3DinAction pipeline that first estimates patches moving in time (t-patches) as a key building block, alongside a hierarchical architecture that learns an informative spatio-temporal representation. We show that our method achieves improved performance on existing datasets, including DFAUST and IKEA ASM.Code is publicly available at https://github.com/sitzikbs/3dincaction

We introduce DyNFL, a novel neural field-based approach for high-fidelity re-simulation of LiDAR scans in dynamic driving scenes. DyNFL processes LiDAR measurements from dynamic environments, accompanied by bounding boxes of moving objects, to construct an editable neural field. This field, comprising separately reconstructed static backgrounds and dynamic objects, allows users to modify viewpoints, adjust object positions, and seamlessly add or remove objects in the re-simulated scene. A key innovation of our method is the neural field composition technique, which effectively integrates reconstructed neural assets from various scenes through a ray drop test, accounting for occlusions and transparent surfaces. Our evaluation with both synthetic and real-world environments demonstrates that DyNFL substantial improves dynamic scene simulation based on LiDAR scans, offering a combination of physical fidelity and flexible editing capabilities.

Inverse rendering, aiming at recovering both the geometry and materials of objects, provides a more compatible reconstruction to conventional rendering engines compared with the popular neural radiance fields (NeRFs). However, existing NeRF-based inverse rendering methods cannot handle glossy objects with local light interactions well, as these methods typically oversimplify the illumination as a 2D environmental map, which assumes infinite lights only. Observing the superiority of NeRFs in recovering radiance fields, we propose a novel 5D Neural Plenoptic Function (NeP) based on NeRFs and ray tracing, such that more accurate lighting-object interactions can be formulated via the rendering equation. We also design a material-aware cone sampling strategy to efficiently integrate lights inside the BRDF lobes with the assistance of pre-filtered radiance fields. Our method is divided into two stages, the geometry of the target object and the pre-filtered environmental radiance fields are reconstructed in the first stage, and materials of the target object are estimated in the second stage with the proposed NeP and material-aware cone sampling strategy. Extensive experiments on the proposed real-world and synthetic datasets demonstrate that our method can reconstruct both high-fidelity geometry and materials of challenging glossy objects with complex lighting interactions from nearby objects. We will release the code and dataset.

Scaled relative pose estimation, i.e., estimating relative rotation and scaled relative translation between two images, has always been a major challenge in global Structure-from-Motion (SfM). This difficulty arises because the two-view relative translation computed by traditional geometric vision methods, e.g. the five-point algorithm, is scaleless. Many researchers have proposed diverse translation averaging methods to solve this problem. Instead of solving the problem in the motion averaging phase, we focus on estimating scaled relative pose with the help of panoramic cameras and deep neural networks. In this paper, a novel network, namely PanoPose, is proposed to estimate the relative motion in a fully self-supervised manner and a global SfM pipeline is built for panorama images. The proposed PanoPose comprises a depth-net and a pose-net, with self-supervision achieved by reconstructing the reference image from its neighboring images based on the estimated depth and relative pose. To maintain precise pose estimation under large viewing angle differences, we randomly rotate the panoramic images and pre-train the pose-net with images before and after the rotation. To enhance scale accuracy, a fusion block is introduced to incorporate depth information into pose estimation. Extensive experiments on panoramic SfM datasets demonstrate the effectiveness of PanoPose compared with state-of-the-arts.

Point clouds captured by different sensors such as RGB-D cameras and LiDAR possess non-negligible domain gaps. Most existing methods design different network architectures and train separately on point clouds from various sensors. Typically, point-based methods achieve outstanding performances on even-distributed dense point clouds from RGB-D cameras, while voxel-based methods are more efficient for large-range sparse LiDAR point clouds. In this paper, we propose geometry-to-occupancy auxiliary learning to enable voxel representations to access point-level geometric information, which supports better generalisation of the voxel-based backbone with additional interpretations of multi-sensor point clouds. Specifically, we construct hierarchical geometry pools generated by a voxel-guided dynamic point network, which efficiently provide auxiliary fine-grained geometric information adapted to different stages of voxel features. We conduct extensive experiments on joint multi-sensor datasets to demonstrate the effectiveness of GeoAuxNet. Enjoying elaborate geometric information, our method outperforms other models collectively trained on multi-sensor datasets, and achieve competitive results with the-state-of-art experts on each single dataset.

This paper targets high-fidelity and real-time view synthesis of dynamic 3D scenes at 4K resolution. Recent methods on dynamic view synthesis have shown impressive rendering quality. However, their speed is still limited when rendering high-resolution images. To overcome this problem, we propose 4K4D, a 4D point cloud representation that supports hardware rasterization and network pre-computation to enable unprecedented rendering speed with a high rendering quality. Our representation is built on a 4D feature grid so that the points are naturally regularized and can be robustly optimized. In addition, we design a novel hybrid appearance model that significantly boosts the rendering quality while preserving efficiency. Moreover, we develop a differentiable depth peeling algorithm to effectively learn the proposed model from RGB videos. Experiments show that our representation can be rendered at over 400 FPS on the DNA-Rendering dataset at 1080p resolution and 80 FPS on the ENeRF-Outdoor dataset at 4K resolution using an RTX 4090 GPU, which is 30$\times$ faster than previous methods and achieves the state-of-the-art rendering quality. Our project page is available at https://zju3dv.github.io/4k4d.

We present Multi-Baseline Radiance Fields (MuRF), a general feed-forward approach to solving sparse view synthesis under multiple different baseline settings (small and large baselines, and different number of input views). To render a target novel view, we discretize the 3D space into planes parallel to the target image plane, and accordingly construct a target view frustum volume. Such a target volume representation is spatially aligned with the target view, which effectively aggregates relevant information from the input views for high-quality rendering. It also facilitates subsequent radiance field regression with a convolutional network thanks to its axis-aligned nature. The 3D context modeled by the convolutional network enables our method to synthesis sharper scene structures than prior works. Our MuRF achieves state-of-the-art performance across multiple different baseline settings and diverse scenarios ranging from simple objects (DTU) to complex indoor and outdoor scenes (RealEstate10K and LLFF). We also show promising zero-shot generalization abilities on the Mip-NeRF 360 dataset, demonstrating the general applicability of MuRF.

Human lives in a 3D world and commonly uses natural language to interact with a 3D scene. Modeling a 3D language field to support open-ended language queries in 3D has gained increasing attention recently. This paper introduces LangSplat, which constructs a 3D language field that enables precise and efficient open-vocabulary querying within 3D spaces. Unlike existing methods that ground CLIP language embeddings in a NeRF model, LangSplat advances the field by utilizing a collection of 3D Gaussians, each encoding language features distilled from CLIP, to represent the language field. By employing a tile-based splatting technique for rendering language features, we circumvent the costly rendering process inherent in NeRF. Instead of directly learning CLIP embeddings, LangSplat first trains a scene-wise language autoencoder and then learns language features on the scene-specific latent space, thereby alleviating substantial memory demands imposed by explicit modeling. Existing methods struggle with imprecise and vague 3D language fields, which fail to discern clear boundaries between objects. We delve into this issue and propose to learn hierarchical semantics using SAM, thereby eliminating the need for extensively querying the language field across various scales and the regularization of DINO features. Extensive experiments on open-vocabulary 3D object localization and semantic segmentation demonstrate that LangSplat significantly outperforms the previous state-of-the-art method LERF. Notably, LangSplat is extremely efficient, achieving a 119 $\times$ speedup compared to LERF.

Neural Radiance Fields (NeRFs) have shown promise in applications like view synthesis and depth estimation, butlearning from multiview images faces inherent uncertainties. Current methods to quantify them are either heuristicor computationally demanding. We introduce BayesRays, a post-hoc framework to evaluate uncertainty in any pretrained NeRF without modifying the training process. Our method establishes a volumetric uncertainty field using spatial perturbations and a Bayesian Laplace approximation. We derive our algorithm statistically and show its superior performance in key metrics and applications. Additional results available at: https://bayesrays.github.io/

We present an approach to accelerate Neural Field training by efficiently selecting sampling locations. While Neural Fields have recently become popular, it is often trained by uniformly sampling the training domain, or through handcrafted heuristics. We show that improved convergence and final training quality can be achieved by a soft mining technique based on importance sampling: rather than either considering or ignoring a pixel completely, we weigh the corresponding loss by a scalar. To implement our idea we use Langevin Monte-Carlo sampling. We show that by doing so, regions with higher error are being selected more frequently, leading to more than 2x improvement in convergence speed.

Novel view synthesis is attractive for social media, but it often contains unwanted details such as personal information that needs to be edited out for a better experience. Multiplane image (MPI) is desirable for social media because of its generality but it is complex and computationally expensive, making object removal challenging. To address these challenges, we propose CORE-MPI, which employs embedding images to improve the consistency and accessibility of MPI object removal. CORE-MPI allows for real-time transmission and interaction with embedding images on social media, facilitating object removal with a single mask. However, recovering the geometric information hidden in the embedding images is a significant challenge. Therefore, we propose a dual-network approach, where one network focuses on color restoration and the other on inpainting the embedding image including geometric information. For the training of CORE-MPI, we introduce a pseudo-reference loss aimed at proficient color recovery, even in complex scenes or with large masks. Furthermore, we present a disparity consistency loss to preserve the geometric consistency of the inpainted region. We demonstrate the effectiveness of CORE-MPI on RealEstate10K and UCSD datasets.

Human avatar has become a novel type of 3D asset with various applications. Ideally, a human avatar should be fully customizable to accommodate different settings and environments. In this work, we introduce NECA, an approach capable of learning versatile human representation from monocular or sparse-view videos, enabling granular customization across aspects such as pose, shadow, shape, lighting and texture. At the core of our approach is to represent humans in complementary dual spaces and predict disentangled neural fields of geometry, albedo, shadow, as well as an external lighting, from which we are able to derive realistic rendering with high-frequency details via volumetric rendering. Extensive experiments demonstrate the advantage of our method over the state-of-the-art methods in photorealistic rendering, as well as various editing tasks such as novel pose synthesis and relighting.

Current 3D stylization methods often assume static scenes, which violates the dynamic nature of our real world. To address this limitation, we present S-DyRF, a reference-based spatio-temporal stylization method for dynamic neural radiance fields. However, stylizing dynamic 3D scenes is inherently challenging due to the limited availability of stylized reference images along the temporal axis. Our key insight lies in introducing additional temporal cues besides the provided reference. To this end, we generate temporal pseudo-references from the given stylized reference. These pseudo-references facilitate the propagation of style information from the reference to the entire dynamic 3D scene. For coarse style transfer, we enforce novel views and times to mimic the style details present in pseudo-references at the feature level. To preserve high-frequency details, we create a collection of stylized temporal pseudo-rays from temporal pseudo-references. These pseudo-rays serve as detailed and explicit stylization guidance for achieving fine style transfer. Experiments on both synthetic and real-world datasets demonstrate that our method yields plausible stylized results of space-time view synthesis on dynamic 3D scenes.

Recently, the rise of query-based Transformer decoders is reshaping camera-based 3D object detection. These query-based decoders are surpassing the traditional dense BEV (Bird's Eye View)-based methods. However, we argue that dense BEV frameworks remain important due to their outstanding abilities in depth estimation and object localization, depicting 3D scenes accurately and comprehensively. This paper aims to address the drawbacks of the existing dense BEV-based 3D object detectors by introducing our proposed enhanced components, including a CRF-modulated depth estimation module enforcing object-level consistencies, a long-term temporal aggregation module with extended receptive fields, and a two-stage object decoder combining perspective techniques with CRF-modulated depth embedding. These enhancements lead to a ``modernized'' dense BEV framework dubbed BEVNeXt. On the nuScenes benchmark, BEVNeXt outperforms both BEV-based and query-based frameworks under various settings, achieving a state-of-the-art result of 64.2 NDS on the nuScenes test set.

Camera-based Semantic Scene Completion (SSC) is to infer the full geometry of objects and scenes from only 2D images. The task is particularly challenging for those invisible areas, due to the inherent occlusions and lighting ambiguity. Existing works ignore the information missing or ambiguous in those shaded and occluded areas, resulting in distorted geometric prediction. To address this issue, we propose a novel method, Bi-SSC, bidirectional geometric semantic fusion for camera-based 3D semantic scene completion. The key insight is to use the neighboring structure of objects in the image and the spatial differences from different perspectives to compensate for the lack of information in occluded areas. Specifically, we introduce a spatial sensory fusion module with multiple association attention to improve semantic correlation in geometric distributions. This module works within single view and across stereo views to achieve global spatial consistency. Experimental results demonstrate that Bi-SSC outperforms state-of-the-art camera-based methods on the SemanticKITTI, particularly excelling in those invisible areas.

Multiple camera view (multi-view) setups have proven useful in many computer vision applications. However, the high computational cost associated with multiple views creates a significant challenge for end devices with limited computational resources. In modern CPU, pipelining breaks a longer job into steps and enables parallelism over sequential steps from multiple jobs. Inspired by this, we study selective view pipelining for efficient multi-view understanding, which breaks computation of multiple views into steps, and only computes the most helpful views/steps in a parallel manner for the best efficiency. To this end, we use reinforcement learning to learn a very light view selection module that analyzes the target object or scenario from initial views and selects the next-best-view for recognition or detection for pipeline computation. Experimental results on multi-view classification and detection tasks show that our approach achieves promising performance while using only 2 or 3 out of $N$ available views, significantly reducing computational costs while maintaining parallelism over GPU through selective view pipelining

We aim to generate fine-grained 3D geometry from large-scale sparse LiDAR scans, abundantly captured by autonomous vehicles (AV). Contrary to prior work on AV scene completion, we aim to extrapolate fine geometry from unlabeled and beyond spatial limits of LiDAR scans, taking a step towards generating realistic, high-resolution simulation-ready 3D street environments. We propose hierarchical Generative Cellular Automata (hGCA), a spatially scalable conditional 3D generative model, which grows geometry recursively with local kernels following GCAs, in a coarse-to-fine manner, equipped with a light-weight planner to induce global consistency. Experiments on synthetic scenes show that hGCA generates plausible scene geometry with higher fidelity and completeness compared to state-of-the-art baselines. Our model generalizes strongly from sim-to-real, qualitatively outperforming baselines on the Waymo-open dataset. We also show anecdotal evidence of the ability to create novel objects from real-world geometric cues even when trained on limited synthetic content.

Existing 3D mesh shape evaluation metrics mainly focus on the overall shape but are usually less sensitive to local details. This makes them inconsistent with human evaluation, as human perception cares about both overall and detailed shape. In this paper, we propose an analytic metric named Spectrum Area Under the Curve Difference (SAUCD) that demonstrates better consistency with human evaluation. To compare the difference between two shapes, we first transform the 3D mesh to the spectrum domain using the discrete Laplace-Beltrami operator and Fourier transform. Then, we calculate the Area Under the Curve (AUC) difference between the two spectrums, so that each frequency band that captures either the overall or detailed shape is equitably considered. Taking human sensitivity across frequency bands into account, we further extend our metric by learning suitable weights for each frequency band which better aligns with human perception. To measure the performance of SAUCD, we build a 3D mesh evaluation dataset called Shape Grading, along with manual annotations from more than 800 subjects. By measuring the correlation between our metric and human evaluation, we demonstrate that SAUCD is well aligned with human evaluation, and outperforms previous 3D mesh metrics.

We introduce a novel approach for adapting deep stereonetworks in a collaborative manner. By building over principles of federated learning, we develop a distributed framework allowing for demanding the optimization process to a number of clients deployed in different environments. This makes it possible, for a deep stereo network running on resourced-constrained devices, to capitalize on the adaptation process carried out by other instances of the same architecture, and thus improve its accuracy in challenging environments even when it cannot carry out adaptation on its own. Experimental results show how federated adaptation performs equivalently to on-device adaptation, and even better when dealing with challenging environments.

This paper proposes Instruct 4D-to-4D that achieves 4D awareness and spatial-temporal consistency for 2D diffusion models to generate high-quality instruction-guided dynamic scene editing results. Traditional applications of 2D diffusion models in dynamic scene editing often result in inconsistency, primarily due to their inherent frame-by-frame editing methodology. Addressing the complexities of extending instruction-guided editing to 4D, our key insight is to treat a 4D scene as a pseudo-3D scene, decoupled into two sub-problems: achieving temporal consistency in video editing and applying these edits to the pseudo-3D scene. Following this, we first enhance the Instruct-Pix2Pix (IP2P) model with an anchor-aware attention module for batch processing and consistent editing. Additionally, we integrate optical flow-guided appearance propagation in a sliding window fashion for more precise frame-to-frame editing and incorporate depth-based projection to manage the extensive data of pseudo-3D scenes, followed by iterative editing to achieve convergence. We extensively evaluate our approach in various scenes and editing instructions, and demonstrate that it achieves spatially and temporally consistent editing results, with significantly enhanced detail and sharpness over the prior art. Notably, Instruct 4D-to-4D is general and applicable to both monocular and challenging multi-camera scenes. Code and more results are available at immortalco.github.io/Instruct-4D-to-4D.

We propose a novel framework for real-time acquisition and reconstruction of temporally-varying 3D phenomena with high quality. The core of our framework is a deep neural network, with an encoder that directly maps to the structured illumination during acquisition, a decoder that predicts a 1D density distribution from single-pixel measurements under the optimized lighting, and an aggregation module that combines the predicted densities for each camera into a single volume. It enables the automatic and joint optimization of physical acquisition and computational reconstruction, and is flexible to adapt to different hardware configurations. The effectiveness of our framework is demonstrated on a lightweight setup with an off-the-shelf projector and one or multiple cameras, achieving a performance of $40$ volumes per second at a spatial resolution of $128^3$. We compare favorably with state-of-the-art techniques in real and synthetic experiments, and evaluate the impact of various factors over our pipeline.

This work delves into the task of pose-free novel view synthesis from stereo pairs, a challenging and pioneering task in 3D vision. Our innovative framework, unlike any before, seamlessly integrates 2D correspondence matching, camera pose estimation, and NeRF rendering, fostering a synergistic enhancement of these tasks. We achieve this through designing an architecture that utilizes a shared representation, which serves as a foundation for enhanced 3D geometry understanding. Capitalizing on the inherent interplay between the tasks, our unified framework is trained end-to-end with the proposed training strategy to improve overall model accuracy. Through extensive evaluations across diverse indoor and outdoor scenes from two real-world datasets, we demonstrate that our approach achieves substantial improvement over previous methodologies, especially in scenarios characterized by extreme viewpoint changes and the absence of accurate camera poses.

Matching cost aggregation plays a fundamental role in learning-based multi-view stereo networks. However, directly aggregating adjacent costs can lead to suboptimal results due to local geometric inconsistency. Related methods either seek selective aggregation or improve aggregated depth in the 2D space, both are unable to handle geometric inconsistency in the cost volume effectively. In this paper, we propose GoMVS to aggregate geometrically consistent costs, yielding better utilization of adjacent geometries. More specifically, we correspond and propagate adjacent costs to the reference pixel by leveraging the local geometric smoothness in conjunction with surface normals. We achieve this by the geometric consistent propagation (GCP) module. It computes the correspondence from the adjacent depth hypothesis space to the reference depth space using surface normals, then uses the correspondence to propagate adjacent costs to the reference geometry, followed by a convolution for aggregation. Our method achieves new state-of-the-art performance on DTU, Tanks & Temple, and ETH3D datasets. Notably, our method ranks 1st on the Tanks & Temple Advanced benchmark. Code is available at https://github.com/Wuuu3511/GoMVS.

Feature matching is a crucial task in the field of computer vision, which involves finding correspondences between images. Previous studies achieve remarkable performance using learning-based feature comparison. However, the pervasive presence of matching redundancy between images gives rise to unnecessary and error-prone computations in these methods, imposing limitations on their accuracy. To address this issue, we propose MESA, a novel approach to establish precise area (or region) matches for efficient matching redundancy reduction. MESA first leverages the advanced image understanding capability of SAM, a state-of-the-art foundation model for image segmentation, to obtain image areas with implicit semantic. Then, a multi-relational graph is proposed to model the spatial structure of these areas and construct their scale hierarchy. Based on graphical models derived from the graph, the area matching is reformulated as an energy minimization task and effectively resolved. Extensive experiments demonstrate that MESA yields substantial precision improvement for multiple point matchers in indoor and outdoor downstream tasks, e.g. +13.61% for DKM in indoor pose estimation.

We present a method to reconstruct indoor and outdoor static scene geometry and appearance from an omnidirectional video moving in a small circular sweep. This setting is challenging because of the small baseline and large depth ranges. These create large variance in the estimation of ray crossings, and make optimization of the surface geometry challenging. To better constrain the optimization, we estimate the geometry as a signed distance field within a spherical binoctree data structure, and use a complementary efficient tree traversal strategy based on breadth-first search for sampling. Unlike regular grids or trees, the shape of this structure well-matches the input camera setting, creating a better trade-off in the memory-quality-compute space. Further, from an initial dense depth estimate, the binoctree is adaptively subdivided throughout optimization. This is different from previous methods that may use a fixed depth, leaving the scene undersampled. In comparisons with three current methods (one neural optimization and two non-neural), our method shows decreased geometry error on average, especially in a detailed scene, while requiring orders of magnitude fewer cells than naive grids for the same minimum voxel size.

Nowadays, leveraging 2D images and pre-trained models to guide 3D point cloud feature representation has shown a remarkable potential to boost the performance of 3D fundamental models. While some works rely on additional data such as 2D real-world images and their corresponding camera poses, recent studies target at using point cloud exclusively by designing 3D-to-2D projection. However, in the indoor scene scenario, existing 3D-to-2D projection strategies suffer from severe occlusions and incoherence, which fail to contain sufficient information for fine-grained point cloud segmentation task. In this paper, we argue that the crux of the matter resides in the basic premise of existing projection strategies that the medium is homogeneous, thereby projection rays propagate along straight lines and behind objects are occluded by front ones. Inspired by the phenomenon of mirage where the occluded objects are exposed by distorted light rays due to heterogeneous medium refraction rate, we propose MirageRoom by designing parametric mirage projection with heterogeneous medium to obtain series of projected images with various distorted degrees. We further develop a masked reprojection module across 2D and 3D latent space to bridge the gap between pre-trained 2D backbone and 3D point-wise features. Both quantitative and qualitative experimental results on S3DIS and ScanNet V2 demonstrate the effectiveness of our method.

With advancements in domain generalized stereo matching networks, models pre-trained on synthetic data demonstrate strong robustness to unseen domains. However, few studies have investigated the robustness after fine-tuning them in real-world scenarios, during which the domain generalization ability can be seriously degraded. In this paper, we explore fine-tuning stereo matching networks without compromising their robustness to unseen domains. Our motivation stems from comparing Ground Truth (GT) versus Pseudo Label (PL) for fine-tuning: GT degrades, but PL preserves the domain generalization ability. Empirically, we find the difference between GT and PL implies valuable information that can regularize networks during fine-tuning. We also propose a framework to utilize this difference for fine-tuning, consisting of a frozen Teacher, an exponential moving average (EMA) Teacher, and a Student network. The core idea is to utilize the EMA Teacher to measure what the Student has learned and dynamically improve GT and PL for fine-tuning. We integrate our framework with state-of-the-art networks and evaluate its effectiveness on several real-world datasets. Extensive experiments show that our method effectively preserves the domain generalization ability during fine-tuning.

3D Semantic Scene Completion (SSC) has emerged as a nascent and pivotal undertaking in autonomous driving, aiming to predict voxel occupancy within volumetric scenes.However, prevailing methodologies primarily focus on voxel-wise feature aggregation, while neglecting instance semantics and scene context. In this paper, we present a novel paradigm termed Symphonies (Scene-from-Insts), that delves into the integration of instance queries to orchestrate 2D-to-3D reconstruction and 3D scene modeling. Leveraging our proposed Serial Instance-Propagated Attentions, Symphonies dynamically encodes instance-centric semantics, facilitating intricate interactions between image-based and volumetric domains. Simultaneously, Symphonies enables holistic scene comprehension by capturing context through the efficient fusion of instance queries, alleviating geometric ambiguity such as occlusion and perspective errors through contextual scene reasoning. Experimental results demonstrate that Symphonies achieves state-of-the-art performance on challenging benchmarks—SemanticKITTI and SSCBench-KITTI-360, yielding remarkable mIoU scores of 15.04 and 18.58, respectively. These results showcase the paradigm's promising advancements.

Neural implicit surface reconstruction leveraging volume rendering has led to significant advances in multi-view reconstruction. However, results for transparent objects can be very poor, primarily because the rendering function fails to account for the intricate light transport induced by refraction and reflection. In this study, we introduce transparent neural surface refinement (TNSR), a novel surface reconstruction framework that explicitly incorporates physical refraction and reflection tracing. Beginning with an initial, approximate surface, our method employs sphere tracing combined with Snell's law to cast both reflected and refracted rays. Central to our proposal is an innovative differentiable technique devised to allow signals from the photometric evidence to propagate back to the surface model by considering how the surface bends and reflects light rays. This allows us to connect surface refinement with volume rendering, enabling end-to-end optimization solely on multi-view RGB images. In our experiments, TNSR demonstrates significant improvements in novel view synthesis and geometry estimation of transparent objects, without prior knowledge of the refractive index.

Two-view correspondence learning has recently focused on considering the coherence and smoothness of the motion field between an image pair. Dominant schemes include controlling the complexity of the field function with regularization or smoothing the field with local filters, but the former suffers from heavy computational burden, and the latter fails to accommodate discontinuities in the case of large scene disparities. In this paper, inspired by Fourier expansion, we propose a novel network called DeMatch, which decomposes the motion field to retain its main ``low-frequency'' and smooth part. This achieves implicit regularization with lower computational cost and generates piecewise smoothness naturally. Specifically, we first decompose the rough motion field that is contaminated by false matches into several different sub-fields, which are highly smooth and contain the main energy of the original field. Then, with these smooth sub-fields, we recover a cleaner motion field from which correct motion vectors are subsequently derived. We also design a special masked decomposition strategy to further mitigate the negative influence of false matches. All the mentioned processes are finally implemented in a discrete and learnable manner, avoiding the difficulty of calculating real dense fields. Extensive experiments reveal that DeMatch outperforms state-of-the-art methods in multiple tasks and shows promising low computational usage and piecewise smoothness property. The code and trained models are publicly available at https://github.com/SuhZhang/DeMatch.

Neural Radiance Field (NeRF) has achieved superior performance for novel view synthesis by modeling the scene with a Multi-Layer Perception (MLP) and a volume rendering procedure, however, when fewer known views are given (i.e., few-shot view synthesis), the model is prone to overfit the given views. To handle this issue, previous efforts have been made towards leveraging learned priors or introducing additional regularizations. In contrast, in this paper, we for the first time provide an orthogonal method from the perspective of network structure. Given the observation that trivially reducing the number of model parameters alleviates the overfitting issue, but at the cost of missing details, we propose the multi-input MLP (mi-MLP) that incorporates the inputs (i.e., location and viewing direction) of the vanilla MLP into each layer to prevent the overfitting issue without harming detailed synthesis. To further reduce the artifacts, we propose to model the color and volume density separately and present two regularization terms. Extensive experiments on multiple datasets demonstrate that: 1) although the proposed mi-MLP is easy to implement, it is surprisingly effective as it boosts the PSNR of the baseline from $14.73$ to $24.23$. 2) the overall framework achieves state-of-the-art results on a wide range of benchmarks.

We introduce GaussianAvatars, a new method to create photorealistic head avatars that are fully controllable in terms of expression, pose, and viewpoint. The core idea of our method is a dynamic 3D representation based on 3D Gaussian splats that are rigged to a parametric morphable face model. This combination facilitates photorealistic rendering while also allowing for precise animation control via the underlying parametric model, e.g., through expression transfer from a driving sequence of a different person or by manually changing the morphable model parameters. In addition to the geometry of the morphable model itself, we optimize for explicit displacement offsets to obtain a more accurate geometric representation. Each splat location is then parameterized by a local coordinate frame to compensate for inaccuracies. During avatar reconstruction, we jointly optimize for the morphable model parameters and Gaussian splat parameters in an end-to-end fashion. We demonstrate the animation capabilities of our photorealistic avatar in several challenging scenarios. For instance, we show reenactments from a driving video, where our method outperforms existing works by a significant margin.

Representing and rendering dynamic scenes has been an important but challenging task. Especially, to accurately model complex motions, high efficiency is usually hard to guarantee. To achieve real-time dynamic scene rendering while also enjoying high training and storage efficiency, we propose 4D Gaussian Splatting (4D-GS) as a holistic representation for dynamic scenes rather than applying 3D-GS for each individual frame. In 4D-GS, a novel explicit representation containing both 3D Gaussians and 4D neural voxels is proposed. A decomposed neural voxel encoding algorithm inspired by HexPlane is proposed to efficiently build Gaussian features from 4D neural voxels and then a lightweight MLP is applied to predict Gaussian deformations at novel timestamps. Our 4D-GS method achieves real-time rendering under high resolutions, 82 FPS at an 800$\times$800 resolution on an RTX 3090 GPU while maintaining comparable or better quality than previous state-of-the-art methods. More demos and code are available at https://guanjunwu.github.io/4dgs/.

In recent years, Neural Radiance Field (NeRF) has demonstrated remarkable capabilities in representing 3D scenes. To expedite the rendering process, learnable explicit representations have been introduced for combination with implicit NeRF representation, which however results in a large storage space requirement. In this paper, we introduce the Context-based NeRF Compression (CNC) framework, which leverages highly efficient context models to provide a storage-friendly NeRF representation. Specifically, we excavate both level-wise and dimension-wise context dependencies to enable probability prediction for information entropy reduction. Additionally, we exploit hash collision and occupancy grids as strong prior knowledge for better context modeling. To the best of our knowledge, we are the first to construct and exploit context models for NeRF compression. We achieve a size reduction of 100X and 70X with improved fidelity against the baseline Instant-NGP on Synthesic-NeRF and Tanks and Temples datasets, respectively. Additionally, we attain 86.7% and 82.3% storage size reduction against the SOTA NeRF compression method BiRF. Our code will be publicly available.

Implicit neural representation has paved the way for new approaches to dynamic scene reconstruction and rendering. Nonetheless, cutting-edge dynamic neural rendering methods rely heavily on these implicit representations, which frequently struggle to capture the intricate details of objects in the scene. Furthermore, implicit methods have difficulty achieving real-time rendering in general dynamic scenes, limiting their use in a variety of tasks. To address the issues, we propose a deformable 3D Gaussians Splatting method that reconstructs scenes using 3D Gaussians and learns them in canonical space with a deformation field to model monocular dynamic scenes. We also introduce an annealing smoothing training mechanism with no extra overhead, which can mitigate the impact of inaccurate poses on the smoothness of time interpolation tasks in real-world datasets. Through a differential Gaussian rasterizer, the deformable 3D Gaussians not only achieve higher rendering quality but also real-time rendering speed. Experiments show that our method outperforms existing methods significantly in terms of both rendering quality and speed, making it well-suited for tasks such as novel-view synthesis, time interpolation, and real-time rendering.

We propose a voxel-based optimization framework, ReVoRF, for few-shot radiance fields that strategically address the unreliability in pseudo novel view synthesis. Our method pivots on the insight that relative depth relationships within neighboring regions are more reliable than the absolute color values in disoccluded areas. Consequently, we devise a bilateral geometric consistency loss that carefully navigates the trade-off between color fidelity and geometric accuracy in the context of depth consistency for uncertain regions. Moreover, we present a reliability-guided learning strategy to discern and utilize the variable quality across synthesized views, complemented by a reliability-aware voxel smoothing algorithm that smoothens the transition between reliable and unreliable data patches. Our approach allows for a more nuanced use of all available data, promoting enhanced learning from regions previously considered unsuitable for high-quality reconstruction. Extensive experiments across diverse datasets reveal that our approach attains significant gains in efficiency and accuracy, delivering rendering speeds of 3 FPS, 7 mins to train a $360^\circ$ scene, and a 5\% improvement in PSNR over existing few-shot methods. Code is available at https://github.com/HKCLynn/ReVoRF.

Neural 3D reconstruction from multi-view images has recently attracted increasing attention from the community. Existing methods normally learn a neural field for the whole scene, while it is still under-explored how to reconstruct a target object indicated by users. Considering the Segment Anything Model (SAM) has shown effectiveness in segmenting any 2D images, in this paper, we propose NTO3D, a novel high-quality Neural Target Object 3D (NTO3D) reconstruction method, which leverages the benefits of both neural field and SAM. We first propose a novel strategy to lift the multi-view 2D segmentation masks of SAM into a unified 3D occupancy field. The 3D occupancy field is then projected into 2D space and generates the new prompts for SAM. This process is iterative until convergence to separate the target object from the scene. After this, we then lift the 2D features of the SAM encoder into a 3D feature field in order to improve the reconstruction quality of the target object. NTO3D lifts the 2D masks and features of SAM into the 3D neural field for high-quality neural target object 3D reconstruction. We conduct detailed experiments on several benchmark datasets to demonstrate the advantages of our method. The code will be available at: https://github.com/ucwxb/NTO3D.

Neural RGBD SLAM techniques have shown promise in dense Simultaneous Localization And Mapping (SLAM), yet face challenges such as error accumulation during camera tracking resulting in distorted maps. In response, we introduce Loopy-SLAM that globally optimizes poses and the dense 3D model. We use frame to model tracking using a data-driven point-based submap generation method and trigger loop closures online by performing global place recognition. Robust pose graph optimization is used to rigidly align the local submaps. As our representation is point based, map corrections can be performed efficiently without the need to store the entire history of input frames as required by methods employing a grid based mapping structure. Evaluation on the synthetic Replica and real-world TUM-RGBD and ScanNet datasets demonstrate competitive or superior performance in tracking, mapping, and rendering accuracy when compared to existing dense neural RGBD SLAM methods. Our source code will be made available.

3D instance segmentation (3DIS) is a crucial task, but point-level annotations are tedious in fully supervised settings. Thus, using bounding boxes (bboxes) as annotations has shown great potential. The current mainstream approach is a two-step process, involving the generation of pseudo-labels from box annotations and the training of a 3DIS network with the pseudo-labels. However, due to the presence of intersections among bboxes, not every point has a determined instance label, especially in overlapping areas. To generate higher quality pseudo-labels and achieve more precise weakly supervised 3DIS results, we propose the Box-Supervised Simulation-assisted Mean Teacher for 3D Instance Segmentation (BSNet), which devises a novel pseudo-labeler called Simulation-assisted Transformer. The labeler consists of two main components. The first is Simulation-assisted Mean Teacher, which introduces Mean Teacher for the first time in this task and constructs simulated samples to assist the labeler in acquiring prior knowledge about overlapping areas. To better model local-global structure, we also propose Local-Global Aware Attention as the decoder for teacher and student labelers. Extensive experiments conducted on the ScanNetV2 and S3DIS datasets verify the superiority of our designs.

We propose ExtraNeRF, a novel method for extrapolating the range of views handled by a Neural Radiance Field (NeRF). Our main idea is to leverage NeRFs to model scene-specific, fine-grained details, while capitalizing on diffusion models to extrapolate beyond our observed data. A key ingredient is to track visibility to determine what portions of the scene have not been observed, and focus on reconstructing those regions consistently with diffusion models. Our primary contributions include a visibility-aware diffusion-based inpainting module that is fine-tuned on the input imagery, yielding an initial NeRF with moderate quality (often blurry) inpainted regions, followed by a second diffusion model trained on the input imagery to consistently enhance, notably sharpen, the inpainted imagery from the first pass. We demonstrate high-quality results, extrapolating beyond the a small number of (typically six or fewer) input views, effectively outpainting the NeRF as well as inpainting newly disoccluded regions inside the original viewing volume. We compare with related work both quantitatively and qualitatively and show significant gains over prior art.

Scale-ambiguity in 3D scene dimensions leads to magnitude-ambiguity of volumetric densities in neural radiance fields, i.e., the densities double when scene size is halved, and vice versa. We call this property alpha invariance. For NeRFs to better maintain alpha invariance, we recommend 1) parameterizing both distance and volume densities in log space, and 2) a discretization-agnostic initialization strategy to guarantee high ray transmittance. We revisit a few popular radiance field models and find that these systems use various heuristics to deal with issues arising from scene scaling. We test their behaviors and show our recipe to be more robust.

Recovering dense and long-range pixel motion in videos is a challenging problem. Part of the difficulty arises from the 3D-to-2D projection process, leading to occlusions and discontinuities in the 2D motion domain. While 2D motion can be intricate, we posit that the underlying 3D motion can often be simple and low-dimensional. In this work, we propose to estimate point trajectories in 3D space to mitigate the issues caused by image projection. Our method, named SpatialTracker, lifts 2D pixels to 3D using monocular depth estimators, represents the 3D content of each frame efficiently using a triplane representation, and performs iterative updates using a transformer to estimate 3D trajectories. Tracking in 3D allows us to leverage as-rigid-as possible(ARAP) constraints while simultaneously learning a rigidity embedding that clusters pixels into different rigid parts. Extensive evaluation shows that our approach achieves state-of-the-art tracking performance both qualitatively and quantitatively, particularly in chal- lenging scenarios such as out-of-plane rotation. And our project page is available at https://henry123-boy.github.io/SpaTracker/.

We present, GauHuman, a 3D human model with Gaussian Splatting for both fast training (1~2 minutes) and real-time rendering (up to 189 FPS), compared with existing NeRF-based implicit representation modelling frameworks demanding hours of training and seconds of rendering per frame. Specifically, GauHuman encodes Gaussian Splatting in the canonical space and transforms 3D Gaussians from canonical space to posed space with linear blend skinning (LBS), in which effective pose and LBS refinement modules are designed to learn fine details of 3D humans under negligible computational cost. Moreover, to enable fast optimization of GauHuman, we initialize and prune 3D Gaussians with 3D human prior, while splitting/cloning via KL divergence guidance, along with a novel merge operation for further speeding up. Extensive experiments on ZJU_Mocap and MonoCap datasets demonstrate that GauHuman achieves state-of-the-art performance quantitatively and qualitatively with fast training and real-time rendering speed. Notably, without sacrificing rendering quality, GauHuman can fast model the 3D human performer with ~13k 3D Gaussians. Our code is available at https://github.com/skhu101/GauHuman.

Generalizable 3D object reconstruction from single-view RGB-D images remains a challenging task, particularly with real-world data. Current state-of-the-art methods develop Transformer-based implicit field learning, necessitating an intensive learning paradigm that requires dense query-supervision uniformly sampled throughout the entire space. We propose a novel approach, IPoD, which harmonizes implicit field learning with point diffusion. This approach treats the query points for implicit field learning as a noisy point cloud for iterative denoising, allowing for their dynamic adaptation to the target object shape. Such adaptive query points harness diffusion learning's capability for coarse shape recovery and also enhances the implicit representation's ability to delineate finer details. Besides, an additional self-conditioning mechanism is designed to use implicit predictions as the guidance of diffusion learning, leading to a cooperative system. Experiments conducted on the CO3D-v2 dataset affirm the superiority of IPoD, achieving 7.8% improvement in F-score and 28.6% in Chamfer distance over existing methods. The generalizability of IPoD is also demonstrated on the MVImgNet dataset. Our project page is at https://yushuang-wu.github.io/IPoD.

Recent advancements in vision-language foundation models have significantly enhanced open-vocabulary 3D scene understanding. However, the generalizability of existing methods is constrained due to their framework designs and their reliance on 3D data. We address this limitation by introducing Generalizable Open-Vocabulary Neural Semantic Fields (GOV-NeSF), a novel approach offering a generalizable implicit representation of 3D scenes with open-vocabulary semantics. We aggregate the geometry-aware features using a cost volume, and propose a Multi-view Joint Fusion module to aggregate multi-view features through a cross-view attention mechanism, which effectively predicts view-specific blending weights for both colors and open-vocabulary features. Remarkably, our GOV-NeSF exhibits state-of-the-art performance in both 2D and 3D open-vocabulary semantic segmentation, eliminating the need for ground truth semantic labels or depth priors, and effectively generalize across scenes and datasets without fine-tuning.

Instance shape reconstruction from a 3D scene involves recovering the full geometries of multiple objects at the semantic instance level. Many methods leverage data-driven learning due to the intricacies of scene complexity and significant indoor occlusions. Training these methods often requires a large-scale, high-quality dataset with aligned and paired shape annotations with real-world scans. Existing datasets are either synthetic or misaligned, restricting the performance of data-driven methods on real data. To this end, we introduce LASA, a Large-scale Aligned Shape Annotation Dataset comprising 10,412 high-quality CAD annotations aligned with 920 real-world scene scans from ArkitScenes, created manually by professional artists. On this top, we propose a novel Diffusion-based Cross-Modal Shape Reconstruction (DisCo) method. It is empowered by a hybrid feature aggregation design to fuse multi-modal inputs and recover high-fidelity object geometries. Besides, we present an Occupancy-Guided 3D Object Detection (OccGOD) method and demonstrate that our shape annotations provide scene occupancy clues that can further improve 3D object detection. Supported by LASA, extensive experiments show that our methods achieve state-of-the-art performance in both instance-level scene reconstruction and 3D object detection tasks.

Can we synthesize 3D humans interacting with scenes without learning from any 3D human-scene interaction data? We propose GenZI, the first zero-shot approach to generating 3D human-scene interactions. Key to GenZI is our distillation of interaction priors from large vision-language models (VLMs), which have learned a rich semantic space of 2D human-scene compositions. Given a natural language description and a coarse point location of the desired interaction in a 3D scene, we first leverage VLMs to imagine plausible 2D human interactions inpainted into multiple rendered views of the scene. We then formulate a robust iterative optimization to synthesize the pose and shape of a 3D human model in the scene, guided by consistency with the 2D interaction hypotheses. In contrast to existing learning-based approaches, GenZI circumvents the conventional need for captured 3D interaction data, and allows for flexible control of the 3D interaction synthesis with easy-to-use text prompts. Extensive experiments show that our zero-shot approach has high flexibility and generality, making it applicable to diverse scene types, including both indoor and outdoor environments.

Multi-view photometric stereo (MVPS) recovers a high-fidelity 3D shape of a scene by benefiting from both multi-view stereo and photometric stereo. While photometric stereo boosts detailed shape reconstruction, it necessitates recording images under various light conditions for each viewpoint. In particular, calibrating the light directions for each view significantly increases the cost of acquiring images. To make MVPS more accessible, we introduce a practical and easy-to-implement setup, multi-view constrained photometric stereo (MVCPS), where the light directions are unknown but constrained to move together with the camera. Unlike conventional multi-view uncalibrated photometric stereo, our constrained setting reduces the ambiguities of surface normal estimates from per-view linear ambiguities to a single and global linear one, thereby simplifying the disambiguation process. The proposed method integrates the ambiguous surface normal into neural surface reconstruction (NeuS) to simultaneously resolve the global ambiguity and estimate the detailed 3D shape. Experiments demonstrate that our method estimates accurate shapes under sparse viewpoints using only a few multi-view constrained light sources.

Determining the relative pose of an object between two images is pivotal to the success of generalizable object pose estimation. Existing approaches typically approximate the continuous pose representation with a large number of discrete pose hypotheses, which incurs a computationally expensive process of scoring each hypothesis at test time. By contrast, we present a Deep Voxel Matching Network (DVMNet) that eliminates the need for pose hypotheses and computes the relative object pose in a single pass. To this end, we map the two input RGB images, reference and query, to their respective voxelized 3D representations. We then pass the resulting voxels through a pose estimation module, where the voxels are aligned and the pose is computed in an end-to-end fashion by solving a least-squares problem. To enhance robustness, we introduce a weighted closest voxel algorithm capable of mitigating the impact of noisy voxels. We conduct extensive experiments on the CO3D, LINEMOD, and Objaverse datasets, demonstrating that our method delivers more accurate relative pose estimates for novel objects at a lower computational cost compared to state-of-the-art methods.

We introduce Motion2VecSets, a 4D diffusion model for dynamic surface reconstruction from point cloud sequences. While existing state-of-the-art methods have demonstrated success in reconstructing non-rigid objects using neural field representations, conventional feed-forward networks encounter challenges with ambiguous observations from noisy, partial, or sparse point clouds. To address these challenges, we introduce a diffusion model that explicitly learns the shape and motion distribution of non-rigid objects through an iterative denoising process of compressed latent representations. The diffusion-based priors enable more plausible and probabilistic reconstructions when handling ambiguous inputs. We parameterize 4D dynamics with latent sets instead of using global latent codes. This novel 4D representation allows us to learn local shape and deformation patterns, leading to more accurate non-linear motion capture and significantly improving generalizability to unseen motions and identities. For more temporally-coherent object tracking, we synchronously denoise deformation latent sets and exchange information across multiple frames. To avoid computational overhead, we designed an interleaved space and time attention block to alternately aggregate deformation latents along spatial and temporal domains. Extensive comparisons against state-of-the-art methods demonstrate the superiority of our Motion2VecSets in 4D reconstruction from various imperfect observations.

We present DiffuScene for indoor 3D scene synthesis based on a novel scene configuration denoising diffusion model. It generates 3D instance properties stored in an unordered object set and retrieves the most similar geometry for each object configuration, which is characterized as a concatenation of different attributes, including location, size, orientation, semantics, and geometry features. We introduce a diffusion network to synthesize a collection of 3D indoor objects by denoising a set of unordered object attributes. Unordered parametrization simplifies and eases the joint distribution approximation. The shape feature diffusion facilitates natural object placements, including symmetries. Our method enables many downstream applications, including scene completion, scene arrangement, and text-conditioned scene synthesis. Experiments on the 3D-FRONT dataset show that our method can synthesize more physically plausible and diverse indoor scenes than state-of-the-art methods. Extensive ablation studies verify the effectiveness of our design choice in scene diffusion models.

There exists an abundance of off-the-shelf models, pretrained on some curated datasets; when tested on new unseen datasets, their performance degrades due to a domain gap between the source training and target testing data. Existing methods for bridging this gap, such as domain adaptation (DA), may require the source data on which the model was trained (often not available), while others, i.e., source-free DA, require many passes through the testing data. We propose an online test-time adaptation method for depth completion, the task of inferring a dense depth map from a single image and associated sparse depth map, that closes the performance gap in a single pass. We first present a study on how the domain shift in each data modality affects model performance. Based on our observations that the sparse depth modality exhibits a much smaller covariate shift than the image, we design an embedding module trained in the source domain that preserves a mapping from features encoding only sparse depth to those encoding image and sparse depth. During test time, sparse depth features are projected using this map as a proxy for source domain features and are used as guidance to train a set of auxiliary parameters (i.e. meta layer) to align image and sparse depth features from the target test domain to that of the source domain. We evaluate our method on indoor and outdoor scenarios and show that it improves over baselines by an average of 21.09\%.

It is challenging for Neural Radiance Fields (NeRFs) in the few-shot setting to reconstruct high-quality novel views and depth maps in $360^\circ$ outward-facing indoor scenes. The captured sparse views for these scenes usually contain large viewpoint variations. This greatly reduces the potential consistency between views, leading NeRFs to degrade a lot in these scenarios. Existing methods usually leverage pretrained depth prediction models to improve NeRFs. However, these methods cannot guarantee geometry consistency due to the inherent geometry ambiguity in the pretrained models, thus limiting NeRFs' performance. In this work, we present P\textsuperscript{2}NeRF to capture global and hierarchical geometry consistency priors from pretrained models, thus facilitating few-shot NeRFs in $360^\circ$ outward-facing indoor scenes. On the one hand, we propose a matching-based geometry warm-up strategy to provide global geometry consistency priors for NeRFs. This effectively avoids the overfitting of early training with sparse inputs. On the other hand, we propose a group depth ranking loss and ray weight mask regularization based on the monocular depth estimation model. This provides hierarchical geometry consistency priors for NeRFs. As a result, our approach can fully leverage the geometry consistency priors from pretrained models and help few-shot NeRFs achieve state-of-the-art performance on two challenging indoor datasets. Our code is released at https://github.com/XT5un/P2NeRF.

In this paper, we present KP-RED, a unified KeyPoint-driven REtrieval and Deformation framework that takes object scans as input and jointly retrieves and deforms the most geometrically similar CAD models from a pre-processed database to tightly match the target.Unlike existing dense matching based methods that typically struggle with noisy partial scans, we propose to leverage category-consistent sparse keypoints to naturally handle both full and partial object scans. Specifically, we first employ a lightweight retrieval module to establish a keypoint-based embedding space, measuring the similarity among objects by dynamically aggregating deformation-aware local-global features around extracted keypoints.Objects that are close in the embedding space are considered similar in geometry.Then we introduce the neural cage-based deformation module that estimates the influence vector of each keypoint upon cage vertices inside its local support region to control the deformation of the retrieved shape.Extensive experiments on the synthetic dataset PartNet and the real-world dataset Scan2CAD demonstratethat KP-RED surpasses existing state-of-the-art approaches by a large margin.Codes and trained models will be released soon.

Neural implicit representation of geometric shapes has witnessed considerable advancements in recent years. However, common distance field based implicit representations, specifically signed distance field (SDF) for watertight shapes or unsigned distance field (UDF) for arbitrary shapes, routinely suffer from degradation of reconstruction accuracy when converting to explicit surface points and meshes. In this paper, we introduce a novel neural implicit representation based on unsigned orthogonal distance fields (UODFs). In UODFs, the minimal unsigned distance from any spatial point to the shape surface is defined solely in one orthogonal direction, contrasting with the multi-directional determination made by SDF and UDF. Consequently, every point in the 3D UODFs can directly access its closest surface points along three orthogonal directions. This distinctive feature leverages the accurate reconstruction of surface points without interpolation errors. We verify the effectiveness of UODFs through a range of reconstruction examples, extending from simple watertight or non-watertight shapes to complex shapes that include hollows, internal or assembling structures.

We present DiSR-NeRF, a diffusion-guided framework for view-consistent super-resolution (SR) NeRF. Unlike prior works, we circumvent the requirement for high-resolution (HR) reference images by leveraging existing powerful 2D super-resolution models. Nonetheless, independent SR 2D images are often inconsistent across different views. We thus propose Iterative 3D Synchronization (I3DS) to mitigate the inconsistency problem via the inherent multi-view consistency property of NeRF. Specifically, our I3DS alternates between upscaling low-resolution (LR) rendered images with diffusion models, and updating the underlying 3D representation with standard NeRF training. We further introduce Renoised Score Distillation (RSD), a novel score-distillation objective for 2D image resolution. Our RSD combines features from ancestral sampling and Score Distillation Sampling (SDS) to generate sharp images that are also LR-consistent. Qualitative and quantitative results on both synthetic and real-world datasets demonstrate that our DiSR-NeRF can achieve better results on NeRF super-resolution compared with existing works. Our source-code will be open-source upon paper acceptance.

Largely due to their implicit nature, neural fields lack a direct mechanism for filtering, as Fourier analysis from discrete signal processing is not directly applicable to these representations. Effective filtering of neural fields is critical to enable level-of-detail processing in downstream applications, and support operations that involve sampling the field on regular grids (e.g. marching cubes). Existing methods that attempt to decompose neural fields in the frequency domain either resort to heuristics, or require extensive modifications to the neural field architecture. We show that via a simple modification, one can obtain neural fields that are low-pass filtered, and in turn show how this can be exploited to obtain a frequency decomposition of the entire signal. We demonstrate the validity of our technique by investigating level-of-detail reconstruction, and showing how coarser representations can be computed effectively.

We present SuperNormal, a fast, high-fidelity approach to multi-view 3D reconstruction using surface normal maps. With a few minutes, SuperNormal produces detailed surfaces on par with 3D scanners. We harness volume rendering to optimize a neural signed distance function (SDF) powered by multi-resolution hash encoding. To accelerate training, we propose directional finite difference and patch-based ray marching to approximate the SDF gradients numerically. While not compromising reconstruction quality, this strategy is nearly twice as efficient as analytical gradients and about three times faster than axis-aligned finite difference. Experiments on the benchmark dataset demonstrate the superiority of SuperNormal in efficiency and accuracy compared to existing multi-view photometric stereo methods. On our captured objects, SuperNormal produces more fine-grained geometry than recent neural 3D reconstruction methods.

A significant challenge facing current optical flow methods is the difficulty in generalizing them well to the real world. This is mainly due to the lack of large-scale real-world datasets, and existing self-supervised methods are limited by indirect loss and occlusions, resulting in fuzzy outcomes.To address this challenge, we introduce a novel optical flow training framework: automatic data factory (ADF). ADF only requires RGB images as input to effectively train the optical flow network on the target data domain.Specifically, we use advanced Nerf technology to reconstruct scenes from photo groups collected by a monocular camera, and then calculate optical flow labels between camera pose pairs based on the rendering results.To eliminate erroneous labels caused by defects in the scene reconstructed by Nerf, we screened the generated labels from multiple aspects, such as optical flow matching accuracy, radiation field confidence, and depth consistency. The filtered labels can be directly used for network supervision.Experimentally, the generalization ability of ADF on KITTI surpasses existing self-supervised optical flow and monocular scene flow algorithms. In addition, ADF achieves impressive results in real-world zero-point generalization evaluations and surpasses most supervised methods.

In the film and gaming industries, achieving a realistic hair appearance typically involves the use of strands originating from the scalp.However, reconstructing these strands from observed surface images of hair presents significant challenges.The difficulty in acquiring Ground Truth (GT) data has led state-of-the-art learning-based methods to rely on pre-training with manually prepared synthetic CG data.This process is not only labor-intensive and costly but also introduces complications due to the domain gap when compared to real-world data.In this study, we propose an optimization-based approach that eliminates the need for pre-training.Our method represents hair strands as line segments growing from the scalp and optimizes them using a novel differentiable rendering algorithm.To robustly optimize a substantial number of slender explicit geometries, we introduce 3D orientation estimation utilizing global optimization, strand initialization based on Laplace's equation, and reparameterization that leverages geometric connectivity and spatial proximity.Unlike existing optimization-based methods, our method is capable of reconstructing internal hair flow in an absolute direction.Our method exhibits robust and accurate inverse rendering, surpassing the quality of existing methods and significantly improving processing speed.

Towards holistic understanding of 3D scenes, a general 3D segmentation method is needed that can segment diverse objects without restrictions on object quantity or categories, while also reflecting the inherent hierarchical structure. To achieve this, we propose OmniSeg3D, an omniversal segmentation method aims for segmenting anything in 3D all at once. The key insight is to lift multi-view inconsistent 2D segmentations into a consistent 3D feature field through a hierarchical contrastive learning framework, which is accomplished by two steps. Firstly, we design a novel hierarchical representation based on category-agnostic 2D segmentations to model the multi-level relationship among pixels. Secondly, image features rendered from the 3D feature field are clustered at different levels, which can be further drawn closer or pushed apart according to the hierarchical relationship between different levels. In tackling the challenges posed by inconsistent 2D segmentations, this framework yields a global consistent 3D feature field, which further enables hierarchical segmentation, multi-object selection, and global discretization. Extensive experiments demonstrate the effectiveness of our method on high-quality 3D segmentation and accurate hierarchical structure understanding. A graphical user interface further facilitates flexible interaction for omniversal 3D segmentation.

3D Visual Grounding (3DVG) aims at localizing 3D object based on textual descriptions. Conventional supervised methods for 3DVG often necessitate extensive annotations and a predefined vocabulary, which can be restrictive. To address this issue, we propose a novel visual programming approach for zero-shot open-vocabulary 3DVG, leveraging the capabilities of large language models (LLMs). Our approach begins with a unique dialog-based method, engaging with LLMs to establish a foundational understanding of zero-shot 3DVG. Building on this, we design a visual program that consists of three types of modules, i.e., view-independent, view-dependent, and functional modules. Furthermore, we develop an innovative language-object correlation module to extend the scope of existing 3D object detectors into open-vocabulary scenarios. Extensive experiments demonstrate that our zero-shot approach can outperform some supervised baselines, marking a significant stride towards effective 3DVG.

Generating realistic hand motion sequences in interaction with objects has gained increasing attention with the growing interest in digital humans. Prior work has illustrated the effectiveness of employing occupancy-based or distance-based virtual sensors to extract hand-object interaction features. Nonetheless, these methods show limited generalizability across object categories, shapes and sizes. We hypothesize that this is due to two reasons: 1) the limited expressiveness of employed virtual sensors, and 2) scarcity of available training data. To tackle this challenge, we introduce a novel joint-centered sensor designed to reason about local object geometry near potential interaction regions. The sensor queries for object surface points in the neighbourhood of each hand joint. As an important step towards mitigating the learning complexity, we transform the points from global frame to hand template frame and use a shared module to process sensor features of each individual joint. This is followed by a spatio-temporal transformer network aimed at capturing correlation among the joints in different dimensions. Moreover, we devise simple heuristic rules to augment the limited training sequences with vast static hand grasping samples. This leads to a broader spectrum of grasping types observed during training, in turn enhancing our model's generalization capability. We evaluate on two public datasets, GRAB and InterCap, where our method shows superiority over baselines both quantitatively and perceptually. We will release our code and model for further research in this field.

While recent 3D instance segmentation approaches show promising results based on transformer architectures, they often fail to correctly identify instances with similar appearances. They also ambiguously determine edges, leading to multiple misclassifications of adjacent edge points. In this work, we introduce a novel framework, called $\textbf{EASE}$, to overcome these challenges and improve the perception of complex 3D instances. We first propose a semantic guidance network to leverage rich semantic knowledge from a language model as intelligent priors, enhancing the functional understanding of real-world instances beyond relying solely on geometrical information. We explicitly instruct the basic instance queries using text embeddings of each instance to learn deep semantic details. Further, we utilize the edge prediction module, encouraging the segmentation network to be edge-aware. We extract voxel-wise edge maps from point features and use them as auxiliary information for learning edge cues. In our extensive experiments on large-scale benchmarks, ScanNetV2, ScanNet200, S3DIS, and STPLS3D, our EASE outperforms existing state-of-the-art models, demonstrating its superior performance.

Neural rendering methods have significantly advanced photo-realistic 3D scene rendering in various academic and industrial applications. The recent 3D Gaussian Splatting method has achieved the state-of-the-art rendering quality and speed combining the benefits of both primitive-based representations and volumetric representations. However, it often leads to heavily redundant Gaussians that try to fit every training view, neglecting the underlying scene geometry. Consequently, the resulting model becomes less robust to significant view changes, texture-less area and lighting effects. We introduce Scaffold-GS, which uses anchor points to distribute local 3D Gaussians, and predicts their attributes on-the-fly based on viewing direction and distance within the view frustum. Anchor growing and pruning strategies are developed based on the importance of neural Gaussians to reliably improve the scene coverage. We show that our method effectively reduces redundant Gaussians while delivering high-quality rendering. We also demonstrates an enhanced capability to accommodate scenes with varying levels-of-detail and view-dependent observations, without sacrificing the rendering speed. Project page: https://city-super.github.io/scaffold-gs/.

Pose regression networks predict the camera pose of a query image relative to a known environment.Within this family of methods, absolute pose regression (APR) has recently shown promising accuracy in the range of a few centimeters in position error. APR networks encode the scene geometry implicitly in their weights. To achieve high accuracy, they require vast amounts of training data that, realistically, can only be created using novel view synthesis in a days-long process. This process has to be repeated for each new scene again and again. We present a new approach to pose regression, map-relative pose regression (marepo), that satisfies the data hunger of the pose regression network in a scene-agnostic fashion. We condition the pose regressor on a scene-specific map representation such that its pose predictions are relative to the scene map. This allows us to train the pose regressor across hundreds of scenes to learn the generic relation between a scene-specific map representation and the camera pose. Our map-relative pose regressor can be applied to new map representations immediately or after mere minutes of fine-tuning for the highest accuracy. Our approach outperforms previous pose regression methods by far on two public datasets, indoor and outdoor.

Constructing photo-realistic Free-Viewpoint Videos (FVVs) of dynamic scenes from multi-view videos remains a challenging endeavor. Despite the remarkable advancements achieved by current neural rendering techniques, these methods generally require complete video sequences for offline training and are not capable of real-time rendering. To address these constraints, we introduce 3DGStream, a method designed for efficient FVV streaming of real-world dynamic scenes. Our method achieves fast on-the-fly per-frame reconstruction within 12 seconds and real-time rendering at 200 FPS. Specifically, we utilize 3D Gaussians (3DGs) to represent the scene. Instead of the naïve approach of directly optimizing 3DGs per-frame, we employ a compact Neural Transformation Cache (NTC) to model the translations and rotations of 3DGs, markedly reducing the training time and storage required for each FVV frame. Furthermore, we propose an adaptive 3DG addition strategy to handle emerging objects in dynamic scenes. Experiments demonstrate that 3DGStream achieves competitive performance in terms of rendering speed, image quality, training time, and model storage when compared with state-of-the-art methods.

Global translation estimation is a highly challenging step in the global structure from motion (SfM) algorithm.Many existing methods depend solely on relative translations, leading to inaccuracies in low parallax scenes and degradation under collinear camera motion.While recent approaches aim to address these issues by incorporating feature tracks into objective functions, they are often sensitive to outliers.In this paper, we first revisit global translation estimation methods with feature tracks and categorize them into explicit and implicit methods.Then, we highlight the superiority of the objective function based on the cross-product distance metric and propose a novel explicit global translation estimation framework that integrates both relative translations and feature tracks as input.To enhance the accuracy of input observations, we re-estimate relative translations with the coplanarity constraint of the epipolar plane and propose a simple yet effective strategy to select reliable feature tracks.Finally, the effectiveness of our approach is demonstrated through experiments on urban image sequences and unordered Internet images, showcasing its superior accuracy and robustness compared to many state-of-the-art techniques.

Multi-view stereo reconstruction (MVS) in the wild requires to first estimate the camera intrinsic and extrinsic parameters. These are usually tedious and cumbersome to obtain, yet they are mandatory to triangulate corresponding pixels in 3D space, which is at the core of all best performing MVS algorithms. In this work, we take an opposite stance and introduce DUSt3R, a radically novel paradigm for Dense and Unconstrained Stereo 3D Reconstruction of arbitrary image collections, operating without prior information about camera calibration nor viewpoint poses. We cast the pairwise reconstruction problem as a regression of pointmaps, relaxing the hard constraints of usual projective camera models. We show that this formulation smoothly unifies the monocular and binocular reconstruction cases. In the case where more than two images are provided, we further propose a simple yet effective global alignment strategy that expresses all pairwise pointmaps in a common reference frame. We base our network architecture on standard Transformer encoders and decoders, allowing us to leverage powerful pretrained models. Our formulation directly provides a 3D model of the scene as well as depth information, but interestingly, we can seamlessly recover from it, pixel matches, focal lengths, relative and absolute cameras. Extensive experiments on all these tasks showcase how DUSt3R effectively unifies various 3D vision tasks, setting new performance records on monocular & multi-view depth estimation as well as relative pose estimation. In summary, DUSt3R makes many geometric 3D vision tasks easy. Code and models at https://github.com/naver/dust3r

Existing depth sensors are imperfect and may provide inaccurate depth values in challenging scenarios, such as in the presence of transparent or reflective objects. In this work, we present a general framework that leverages polarization imaging to improve inaccurate depth measurements from various depth sensors. Previous polarization-based depth enhancement methods focus on utilizing pure physics-based formulas for a single sensor. In contrast, our method first adopts a learning-based strategy where a neural network is trained to estimate a dense and complete depth map from polarization data and a sensor depth map from different sensors. To further improve the performance, we propose a Polarization Prompt Fusion Tuning (PPFT) strategy to effectively utilize RGB-based models pre-trained on large-scale datasets, as the size of the polarization dataset is limited to train a strong model from scratch. We conducted extensive experiments on a public dataset, and the results demonstrate that the proposed method performs favorably compared to existing depth enhancement baselines. Code and demos are available at https://lastbasket.github.io/PPFT/.

Point cloud filtering is a fundamental 3D vision task, which aims to remove noise while recovering the underlying clean surfaces. State-of-the-art methods remove noise by moving noisy points along stochastic trajectories to the clean surfaces. These methods often require regularization within the training objective and/or during post-processing, to ensure fidelity. In this paper, we introduce StraightPCF, a new deep learning based method for point cloud filtering. It works by moving noisy points along straight paths, thus reducing discretization errors while ensuring faster convergence to the clean surfaces. We model noisy patches as intermediate states between high noise patch variants and their clean counterparts, and design the VelocityModule to infer a constant flow velocity from the former to the latter. This constant flow leads to straight filtering trajectories. In addition, we introduce a DistanceModule that scales the straight trajectory using an estimated distance scalar to attain convergence near the clean surface. Our network is lightweight and only has $\sim530K$ parameters, being 17\% of IterativePFN (a most recent point cloud filtering network). Extensive experiments on both synthetic and real-world data show our method achieves state-of-the-art results. Our method also demonstrates nice distributions of filtered points without the need for regularization. The implementation code can be found at: https://github.com/ddsediri/StraightPCF.

We propose NeRFiller, an approach that completes missing portions of a 3D capture via generative 3D inpainting using off-the-shelf 2D visual generative models. Often parts of a captured 3D scene or object are missing due to mesh reconstruction failures or a lack of observations (e.g., contact regions, such as the bottom of objects, or hard-to-reach areas). We approach this challenging 3D inpainting problem by leveraging a 2D inpainting diffusion model. We identify a surprising behavior of these models, where they generate more 3D consistent inpaints when images form a 2$\times$2 grid, and show how to generalize this behavior to more than four images. We then present an iterative framework to distill these inpainted regions into a single consistent 3D scene. In contrast to related works, we focus on completing scenes rather than deleting foreground objects, and our approach does not require tight 2D object masks or text. We compare our approach to relevant baselines adapted to our setting on a variety of scenes, where NeRFiller creates the most 3D consistent and plausible scene completions.

Neural Rendering representations have significantly contributed to the field of 3D computer vision. Given their potential, considerable efforts have been invested to improve their performance. Nonetheless, the essential question of selecting training views is yet to be thoroughly investigated. This key aspect plays a vital role in achieving high-quality results and aligns with the well-known tenet of deep learning: "garbage in, garbage out". In this paper, we first illustrate the importance of view selection by demonstrating how a simple rotation of the test views within the most pervasive NeRF dataset can lead to consequential shifts in the performance rankings of state-of-the-art techniques. To address this challenge, we introduce a unified framework for view selection methods and devise a thorough benchmark to assess its impact. Significant improvements can be achieved without leveraging error or uncertainty estimation but focusing on uniform view coverage of the reconstructed object, resulting in a training-free approach. Using this technique, we show that high-quality renderings can be achieved faster by using fewer views. We conduct extensive experiments on both synthetic datasets and realistic data to demonstrate the effectiveness of our proposed method compared with random, conventional error-based, and uncertainty-guided view selection.

Geometric knowledge has been shown to be beneficial for the stereo matching task. However, prior attempts to integrate geometric insights into stereo matching algorithms have largely focused on geometric knowledge from single images while crucial cross-view factors such as occlusion and matching uniqueness have been overlooked. To address this gap, we propose a novel Intra-view and Cross-view Geometric knowledge learning Network (ICGNet), specifically crafted to assimilate both intra-view and cross-view geometric knowledge. ICGNet harnesses the power of interest points to serve as a channel for intra-view geometric understanding. Simultaneously, it employs the correspondences among these points to capture cross-view geometric relationships. This dual incorporation empowers the proposed ICGNet to leverage both intra-view and cross-view geometric knowledge in its learning process, substantially improving its ability to estimate disparities. Our extensive experiments demonstrate the superiority of the ICGNet over contemporary leading models and it ranks $1^{st}$ on the KITTI 2015 benchmark among all published methods.

Recently, 3D content creation from text prompts has demonstrated remarkable progress by utilizing 2D and 3D diffusion models. While 3D diffusion models ensure great multi-view consistency, their ability to generate high-quality and diverse 3D assets is hindered by the limited 3D data. In contrast, 2D diffusion models find a distillation approach that achieves excellent generalization and rich details without any 3D data. However, 2D lifting methods suffer from inherent view-agnostic ambiguity thereby leading to serious multi-face Janus issues, where text prompts fail to provide sufficient guidance to learn coherent 3D results.Instead of retraining a costly viewpoint-aware model, we study how to fully exploit easily accessible coarse 3D knowledge to enhance the prompts and guide 2D lifting optimization for refinement. In this paper, we propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric consistency simultaneously. Specifically, we design a pair of guiding strategies derived from the coarse 3D prior generated by the 3D diffusion model: a structural guidance for geometric fidelity and a semantic guidance for 3D coherence. Employing the two types of guidance, the 2D diffusion model enriches the 3D content with diversified and high-quality results. Extensive experiments show the superiority of our Sherpa3D over the state-of-the-art text-to-3D methods in terms of quality and 3D consistency.

Radiance fields have demonstrated impressive performance in synthesizing novel views from sparse input views, yet prevailing methods suffer from high training costs and slow inference speed. This paper introduces DNGaussian, a depth-regularized framework based on 3D Gaussian radiance fields, offering real-time and high-quality few-shot novel view synthesis at low costs. Our motivation stems from the highly efficient representation and surprising quality of the recent 3D Gaussian Splatting, despite it will encounter a geometry degradation when input views decrease. In the Gaussian radiance fields, we find this degradation in scene geometry primarily lined to the positioning of Gaussian primitives and can be mitigated by depth constraint. Consequently, we propose a Hard and Soft Depth Regularization to restore accurate scene geometry under coarse monocular depth supervision while maintaining a fine-grained color appearance. To further refine detailed geometry reshaping, we introduce Global-Local Depth Normalization, enhancing the focus on small local depth changes. Extensive experiments on LLFF, DTU, and Blender datasets demonstrate that DNGaussian outperforms state-of-the-art methods, achieving comparable or better results with significantly reduced memory cost, a $25\times$ reduction in training time, and over $3000\times$ faster rendering speed.

Point cloud upsampling (PCU) enriches the representation of raw point clouds, significantly improving the performance in downstream tasks such as classification and reconstruction. Most of the existing point cloud upsampling methods focus on sparse point cloud feature extraction and upsampling module design. In a different way, we dive deeper into directly modelling the gradient of data distribution from dense point clouds. In this paper, we proposed a conditional denoising diffusion probabilistic model (DDPM) for point cloud upsampling, called PUDM. Specifically, PUDM treats the sparse point cloud as a condition, and iteratively learns the transformation relationship between the dense point cloud and the noise. Simultaneously, PUDM aligns with a dual mapping paradigm to further improve the discernment of point features. In this context, PUDM enables learning complex geometry details in the ground truth through the dominant features, while avoiding an additional upsampling module design. Furthermore, to generate high-quality arbitrary-scale point clouds during inference, PUDM exploits the prior knowledge of the scale between sparse point clouds and dense point clouds during training by parameterizing a rate factor. Moreover, PUDM exhibits strong noise robustness in experimental results. In the quantitative and qualitative evaluations on PU1K and PUGAN, PUDM significantly outperformed existing methods in terms of Chamfer Distance (CD) and Hausdorff Distance (HD), achieving state of the art (SOTA) performance.

While neural rendering has led to impressive advances in scene reconstruction and novel view synthesis, it relies heavily on accurately pre-computed camera poses. To relax this constraint, multiple efforts have been made to train Neural Radiance Fields (NeRFs) without pre-processed camera poses. However, the implicit representations of NeRFs provide extra challenges to optimize the 3D structure and camera poses at the same time. On the other hand, the recently proposed 3D Gaussian Splatting provides new opportunities given its explicit point cloud representations. This paper leverages both the explicit geometric representation and the continuity of the input video stream to perform novel view synthesis without any SfM preprocessing. We process the input frames in a sequential manner and progressively grow the 3D Gaussians set by taking one input frame at a time, without the need to pre-compute the camera poses. Our method significantly improves over previous approaches in view synthesis and camera pose estimation under large motion changes. Our code will be made publicly available.

Utilizing multi-view inputs to synthesize novel-view images, Neural Radiance Fields (NeRF) have emerged as a popular research topic in 3D vision. In this work, we introduce a Generalizable Semantic Neural Radiance Field (GSNeRF), which uniquely takes image semantics into the synthesis process so that both novel view images and the associated semantic maps can be produced for unseen scenes. Our GSNeRF is composed of two stages: Semantic Geo-Reasoning and Depth-Guided Visual rendering. The former is able to observe multi-view image inputs to extract semantic and geometry features from a scene. Guided by the resulting image geometry information, the latter performs both image and semantic rendering with improved performances. Our experiments not only confirm that GSNeRF performs favorably against prior works on both novel-view image and semantic segmentation synthesis but the effectiveness of our sampling strategy for visual rendering is further verified.

Registration of point clouds collected from a pair of distant vehicles provides a comprehensive and accurate 3D view of the driving scenario, which is vital for driving safety related applications, yet existing literature suffers from the expensive pose label acquisition and the deficiency to generalize to new data distributions. In this paper, we propose EYOC, an unsupervised distant point cloud registration method that adapts to new point cloud distributions on the fly, requiring no global pose labels. The core idea of EYOC is to train a feature extractor in a progressive fashion, where in each round, the feature extractor, trained with near point cloud pairs, can label slightly farther point cloud pairs, enabling self-supervision on such far point cloud pairs. This process continues until the derived extractor can be used to register distant point clouds. Particularly, to enable high-fidelity correspondence label generation, we devise an effective spatial filtering scheme to select the most representative correspondences to register a point cloud pair, and then utilize the aligned point clouds to discover more correct correspondences. Experiments show that EYOC can achieve comparable performance with state-of-the-art supervised methods at a lower training cost. Moreover, it outwits supervised methods regarding generalization performance on new data distributions.

We introduce a lightweight and accurate localization method that only utilizes the geometry of 2D-3D lines.Given a pre-captured 3D map, our approach localizes a panorama image, taking advantage of the holistic $360^\circ$ views.The system mitigates potential privacy breaches or domain discrepancies by avoiding trained or hand-crafted visual descriptors.However, as lines alone can be ambiguous, we express distinctive yet compact spatial contexts from relationships between lines, namely the dominant directions of parallel lines and the intersection between non-parallel lines.The resulting representations are efficient in processing time and memory compared to conventional visual descriptor-based methods.Given the groups of dominant line directions and their intersections, we can further accelerate the search process to test thousands of pose candidates in less than a millisecond without sacrificing accuracy.We empirically show that the proposed 2D-3D matching can localize panoramas for challenging scenes with similar structures, dramatic domain shifts or illumination changes.Our fully geometric approach does not involve extensive parameter tuning or neural network training, making it a practical algorithm that can be readily deployed in the real world.

We introduce a novel framework for multiway point cloud mosaicking (named Wednesday), designed to co-align sets of partially overlapping point clouds -- typically obtained from 3D scanners or moving RGB-D cameras -- into a unified coordinate system. At the core of our approach is ODIN, a learned pairwise registration algorithm that iteratively identifies overlaps and refines attention scores, employing a diffusion-based process for denoising pairwise correlation matrices to enhance matching accuracy. Further steps include constructing a pose graph from all point clouds, performing rotation averaging, a novel robust algorithm for re-estimating translations optimally in terms of consensus maximization and translation optimization. Finally, the point cloud rotations and positions are optimized jointly by a diffusion-based approach. Tested on four diverse, large-scale datasets, our method achieves state-of-the-art pairwise and multiway registration results by a large margin on all benchmarks. Our code and models are available at https://github.com/jinsz/Multiway-Point-Cloud-Mosaicking-with-Diffusion-and-Global-Optimization.

Recently, 3D Gaussian Splatting has demonstrated impressive novel view synthesis results, reaching high fidelity and efficiency. However, strong artifacts can be observed when changing the sampling rate, \eg, by changing focal length or camera distance. We find that the source for this phenomenon can be attributed to the lack of 3D frequency constraints and the usage of a 2D dilation filter. To address this problem, we introduce a 3D smoothing filter to constrains the size of the 3D Gaussian primitives based on the maximal sampling frequency induced by the input views. It eliminates high-frequency artifacts when zooming in. Moreover, replacing 2D dilation with a 2D Mip filter, which simulates a 2D box filter, effectively mitigates aliasing and dilation issues.Our evaluation, including scenarios such a training on single-scale images and testing on multiple scales, validates the effectiveness of our approach.

Generative 3D part assembly involves understanding partrelationships and predicting their 6-DoF poses for assembling a realistic 3D shape. Prior work often focus on thegeometry of individual parts, neglecting part-whole hierarchies of objects. Leveraging two key observations: 1)super-part poses provide strong hints about part poses, and2) predicting super-part poses is easier due to fewer superparts, we propose a part-whole-hierarchy message passingnetwork for efficient 3D part assembly. We first introducesuper-parts by grouping geometrically similar parts withoutany semantic labels. Then we employ a part-whole hierarchical encoder, wherein a super-part encoder predicts latentsuper-part poses based on input parts. Subsequently, wetransform the point cloud using the latent poses, feeding itto the part encoder for aggregating super-part informationand reasoning about part relationships to predict all partposes. In training, only ground-truth part poses are required.During inference, the predicted latent poses of super-partsenhance interpretability. Experimental results on the PartNetdataset show that our method achieves state-of-the-art performance in part and connectivity accuracy and enables aninterpretable hierarchical part assembly. Code is availableat https://github.com/pkudba/3DHPA.

Scene reconstruction from multi-view images is a fundamental problem in computer vision and graphics. Recent neural implicit surface reconstruction methods have achieved high-quality results; however, editing and manipulating the 3D geometry of reconstructed scenes remains challenging due to the absence of naturally decomposed object entities and complex object/background compositions. In this paper, we present Total-Decom, a novel method for decomposed 3D reconstruction with minimal human interaction. Our approach seamlessly integrates the Segment Anything Model (SAM) with hybrid implicit-explicit neural surface representations and a mesh-based region-growing technique for accurate 3D object decomposition. Total-Decom requires minimal human annotations while providing users with real-time control over the granularity and quality of decomposition. We extensively evaluate our method on benchmark datasets and demonstrate its potential for downstream applications, such as animation and scene editing. Codes will be made publicly available.

Keypoints used for image matching often include an estimate of the feature scale and orientation. While recent work has demonstrated the advantages of using feature scales and orientations for relative pose estimation, relatively little work has considered their use for absolute pose estimation. We introduce minimal solutions for absolute pose from two oriented feature correspondences in the general case, or one scaled and oriented correspondence given a known vertical direction. Nowadays, assuming a known direction is not particularly restrictive as modern consumer devices, such as smartphones or drones, are equipped with Inertial Measurement Units (IMU) that provide the gravity direction by default. Compared to traditional absolute pose methods requiring three point correspondences, our solvers need a smaller minimal sample, reducing the cost and complexity of robust estimation. Evaluations on large-scale and public real datasets demonstrate the advantage of our methods for fast and accurate localization in challenging conditions.

Matching 2D keypoints in an image to a sparse 3D point cloud of the scene without requiring visual descriptors has garnered increased interest due to its low memory requirements, inherent privacy preservation, and reduced need for expensive 3D model maintenance compared to visual descriptor-based methods. However, existing algorithms often compromise on performance, resulting in a significant deterioration compared to their descriptor-based counterparts. In this paper, we introduce DGC-GNN, a novel algorithm that employs a global-to-local Graph Neural Network (GNN) that progressively exploits geometric and color cues to represent keypoints, thereby improving matching accuracy. Our procedure encodes both Euclidean and angular relations at a coarse level, forming the geometric embedding to guide the point matching.We evaluate DGC-GNN on both indoor and outdoor datasets, demonstrating that it not only doubles the accuracy of the state-of-the-art visual descriptor-free algorithm but also substantially narrows the performance gap between descriptor-based and descriptor-free methods.

Recent advancements in the study of Neural Radiance Fields (NeRF) for dynamic scenes often involve explicit modeling of scene dynamics. However, this approach faces challenges in modeling scene dynamics in urban environments, where moving objects of various categories and scales are present. In such settings, it becomes crucial to effectively eliminate moving objects to accurately reconstruct static backgrounds. Our research introduces an innovative method, termed here as Entity-NeRF, which combines the strengths of knowledge-based and statistical strategies. This approach utilizes entity-wise statistics, leveraging entity segmentation and stationary entity classification through thing/stuff segmentation. To assess our methodology, we created an urban scene dataset masked with moving objects. Our comprehensive experiments demonstrate that Entity-NeRF notably outperforms existing techniques in removing moving objects and reconstructing static urban backgrounds, both quantitatively and qualitatively.

Recently, impressive results have been achieved in 3D scene editing with text instructions based on a 2D diffusion model. However, current diffusion models primarily generate images by predicting noise in the latent space, and the editing is usually applied to the whole image, which makes it challenging to perform delicate, especially localized, editing for 3D scenes. Inspired by recent 3D Gaussian splatting, we propose a systematic framework, named GaussianEditor, to edit 3D scenes delicately via 3D Gaussians with text instructions. Benefiting from the explicit property of 3D Gaussians, we design a series of techniques to achieve delicate editing. Specifically, we first extract the region of interest (RoI) corresponding to the text instruction, aligning it to 3D Gaussians. The Gaussian RoI is further used to control the editing process. Our framework can achieve more delicate and precise editing of 3D scenes than previous methods while enjoying much faster training speed, i.e. within 20 minutes on a single V100 GPU, more than twice as fast as Instruct-NeRF2NeRF (45 minutes -- 2 hours). The project page is at GaussianEditor.github.io.

Humans can infer 3D structure from 2D images of an object based on past experience and improve their 3D understanding as they see more images. Inspired by this behavior, we introduce SAP3D, a system for 3D reconstruction and novel view synthesis from an arbitrary number of unposed images. Given a few unposed images of an object, we adapt a pre-trained view-conditioned diffusion model together with the camera poses of the images via test-time fine-tuning. The adapted diffusion model and the obtained camera poses are then utilized as instance-specific priors for 3D reconstruction and novel view synthesis. We show that as the number of input images increases, the performance of our approach improves, bridging the gap between optimization-based prior-less 3D reconstruction methods and single-image-to-3D diffusion-based methods. We demonstrate our system on real images as well as standard synthetic benchmarks. Our ablation studies confirm that this adaption behavior is key for more accurate 3D understanding.

3D Gaussians have recently emerged as a highly efficient representation for 3D reconstruction and rendering. Despite its high rendering quality and speed at high resolutions, they both deteriorate drastically when rendered at lower resolutions or from far away camera position. During low resolution or far away rendering, the pixel size of the image can fall below the Nyquist frequency compared to the screen size of each splatted 3D Gaussian and leads to aliasing effect. The rendering is also drastically slowed down by the sequential alpha blending of more splatted Gaussians per pixel. To address these issues, we propose a multi-scale 3D Gaussian splatting algorithm, which maintains Gaussians at different scales to represent the same scene. Higher-resolution images are rendered with more small Gaussians, and lower-resolution images are rendered with fewer larger Gaussians. With similar training time, our algorithm can achieve 13\%-66\% PSNR and 160\%-2400\% rendering speed improvement at 4$\times$-128$\times$ scale rendering on Mip-NeRF360 dataset compared to the single scale 3D Gaussian splatting.

Material appearance is a key component of photorealism, with a pronounced impact on human perception. Although there are many prior works targeting at measuring opaque materials using light-weight setups (e.g., consumer-level cameras), little attention is paid on acquiring the optical properties of translucent materials which are also quite common in nature. In this paper, we present a practical method for acquiring scattering properties of translucent materials, based solely on ordinary images captured with unknown lighting and camera parameters. The key to our method is an inter-pixel translucency prior which states that image pixels of a given homogeneous translucent material typically form curves (dubbed translucent curves) in the RGB space, of which their shapes are determined by the parameters of the material. We leverage this prior in a specially-designed convolutional neural network comprising multiple encoders, a translucency-aware feature fusion module and a cascaded decoder. We demonstrate, through both visual comparisons and quantitative evaluations, that high accuracy can be achieved on a wide range of real-world translucent materials.

Semantic, instance, and panoptic segmentation of 3D point clouds have been addressed using task-specific models of distinct design. Thereby, the similarity of all segmentation tasks and the implicit relationship between them have not been utilized effectively. This paper presents a unified, simple, and effective model addressing all these tasks jointly. The model, named OneFormer3D, performs instance and semantic segmentation consistently, using a group of learnable kernels, where each kernel is responsible for generating a mask for either an instance or a semantic category. These kernels are trained with a transformer-based decoder with unified instance and semantic queries passed as an input. Such a design enables training a model end-to-end in a single run, so that it achieves top performance on all three segmentation tasks simultaneously. Specifically, our OneFormer3D ranks 1st and sets a new state-of-the-art (+2.1 mAP50) in the ScanNet test leaderboard. We also demonstrate the state-of-the-art results in semantic, instance, and panoptic segmentation of ScanNet (+21 PQ), ScanNet200 (+3.8 mAP50), and S3DIS (+0.8 mIoU) datasets.

The pre-training architectures of large language models encompass various types, including autoencoding models, autoregressive models, and encoder-decoder models. We posit that any modality can potentially benefit from a large language model, as long as it undergoes vector quantization to become discrete tokens. Inspired by the General Language Model, we propose a General Point Model (GPM) that seamlessly integrates autoencoding and autoregressive tasks in a point cloud transformer. This model is versatile, allowing fine-tuning for downstream point cloud representation tasks, as well as unconditional and conditional generation tasks. GPM enhances masked prediction in autoencoding through various forms of mask padding tasks, leading to improved performance in point cloud understanding. Additionally, GPM demonstrates highly competitive results in unconditional point cloud generation tasks, even exhibiting the potential for conditional generation tasks by modifying the input's conditional information. Compared to models like Point-BERT, MaskPoint, and PointMAE, our GPM achieves superior performance in point cloud understanding tasks. Furthermore, the integration of autoregressive and autoencoding within the same transformer underscores its versatility across different downstream tasks.

Neural rendering has demonstrated remarkable success in dynamic scene reconstruction. Thanks to the expressiveness of neural representations, prior works can accurately capture the motion and achieve high-fidelity reconstruction of the target object. Despite this, real-world video scenarios often feature large unobserved regions where neural representations struggle to achieve realistic completion. To tackle this challenge, we introduce MorpheuS, a framework for dynamic 360$^{\circ}$ surface reconstruction from a casually captured RGB-D video. Our approach models the target scene as a canonical field that encodes its geometry and appearance, in conjunction with a deformation field that warps points from the current frame to the canonical space. We leverage a view-dependent diffusion prior and distill knowledge from it to achieve realistic completion of unobserved regions. Experimental results on various real-world and synthetic datasets show that our method can achieve high-fidelity 360$^{\circ}$ surface reconstruction of a deformable object from a monocular RGB-D video.

We introduce pixelSplat, a feed-forward model that learns to reconstruct 3D radiance fields parameterized by 3D Gaussian primitives from pairs of images. Our model features real-time and memory-efficient rendering for scalable training as well as fast 3D reconstruction at inference time. To overcome local minima inherent to sparse and locally supported representations, we predict a dense probability distribution over 3D and sample Gaussian means from that probability distribution. We make this sampling operation differentiable via a reparameterization trick, allowing us to back-propagate gradients through the Gaussian splatting representation. We benchmark our method on wide-baseline novel view synthesis on the real-world RealEstate10k and ACID datasets, where we outperform state-of-the-art light field transformers and accelerate rendering by 2.5 orders of magnitude while reconstructing an interpretable and editable 3D radiance field. Additional materials can be found on the anonymous project website (pixelsplat.github.io).

Modeling object dynamics with a neural network is an important problem with numerous applications. Most recent work has been based on graph neural networks. However, physics happens in 3D space, where geometric information potentially plays an important role in modeling physical phenomena. In this work, we propose a novel U-net architecture based on continuous point convolution which naturally embeds information from 3D coordinates and allows for multi-scale feature representations with established downsampling and upsampling procedures. Bottleneck layers in the downsampled point clouds lead to better long-range interaction modeling. Besides, the flexibility of point convolutions allows us to generalize to sparsely sampled points from mesh vertices and dynamically generate features on important interaction points on mesh faces. Experimental results demonstrate that our approach significantly improves the state-of-the-art, especially in scenarios that require accurate gravity or collision reasoning.

Absolute Pose Regression (APR) methods use deep neural networks to directly regress camera poses from RGB images. However, the predominant APR architectures only rely on 2D operations during inference, resulting in limited accuracy of pose estimation due to the lack of 3D geometry constraints or priors. In this work, we propose a test-time refinement pipeline that leverages implicit geometric constraints using a robust feature field to enhance the ability of APR methods to use 3D information during inference. We also introduce a novel Neural Feature Synthesizer (NeFeS) model, which encodes 3D geometric features during training and directly renders dense novel view features at test time to refine APR methods. To enhance the robustness of our model, we introduce a feature fusion module and a progressive training strategy. Our proposed method achieves state-of-the-art single-image APR accuracy on indoor and outdoor datasets.

Neural character models can now reconstruct detailed geometry and texture from video, but they lack explicit shadows and shading, leading to artifacts when generating novel views and poses or during relighting. It is particularly difficult to include shadows as they are a global effect and the required casting of secondary rays is costly. We propose a new shadow model using a Gaussian density proxy thatreplaces sampling with a simple analytic formula. It supports dynamic motion and is tailored for shadow computation, thereby avoiding the affine projection approximation and sorting required by the closely related Gaussian splatting. Combined with a deferred neural rendering model, our Gaussian shadows enable Lambertian shading and shadow casting with minimal overhead. We demonstrate improvedreconstructions, with better separation of albedo, shading, and shadows in challenging outdoor scenes with direct sun light and hard shadows. Our method is able to optimize the light direction without any input from the user. As a result, novel poses have fewer shadow artifacts, and relighting in novel scenes is more realistic compared to the state-of-the-art methods, providing new ways to pose neural characters in novel environments, increasing their applicability. Code available at: https://github.com/LuisBolanos17/GaussianShadowCasting

We propose the problem of point-level 3D scene interpolation, which aims to reconstruct a 3D scene in two different states from multiple views, synthesize a plausible smooth point-level interpolation between the 3D scenes in the two states, and render the 3D scene at any point in time from a novel view, all without any supervision in-between the states. The primary challenge lies in producing a smooth transition between the two states which can exhibit substantial changes in geometry. To tackle it, we leverage recent advances in point renderers, which are naturally suited to representing Lagrangian motion. Our approach works by initially learning a point-based representation of the scene in its starting state, followed by finetuning this model towards the end state. Critical to achieving smooth interpolation of both the scene's geometry and appearance is the choice of the point rendering technique. Different techniques excel along different performance dimensions, and we propose leveraging the recent Proximity Attention Point Rendering (PAPR) technique, which is designed to learn point clouds from scratch and support novel view synthesis of scenes after they undergo non-rigid geometric deformations. Our method, which we dub ``PAPR in Motion'', builds on PAPR's strengths and addresses its weaknesses by developing various regularization techniques specifically for the task. The end result is a method that can effectively bridge large scene changes, producing plausible and smooth geometry and appearance interpolations. When evaluating under diverse motion types, we find that PAPR in Motion achieves superior performance relative to the leading point renderer for dynamic scenes, thereby demonstrating an effective solution to this new and challenging problem of point-level 3D scene interpolation.

In this paper, we present ShapeMatcher, a unified self-supervised learning framework for joint shape canonicalization, segmentation, retrieval and deformation. Given a partially-observed object in an arbitrary pose, we first canonicalize the object by extracting point-wise affine-invariant features, disentangling inherent structure of the object with its pose and size. These learned features are then leveraged to predict semantically consistent part segmentation and corresponding part centers. Next, our lightweight retrieval module aggregates the features within each part as its retrieval token and compare all the tokens with source shapes from a pre-established database to identify the most geometrically similar shape. Finally, we deform the retrieved shape in the deformation module to tightly fit the input object by harnessing part center guided neural cage deformation. The key insight of ShapeMaker is the simultaneous training of the four highly-associated processes: canonicalization, segmentation, retrieval, and deformation, leveraging cross-task consistency losses for mutual supervision. Extensive experiments on synthetic datasets PartNet, ComplementMe, and real-world dataset Scan2CAD demonstrate that ShapeMaker surpasses competitors by a large margin.

We propose XScale-NVS for high-fidelity cross-scale novel view synthesis of real-world large-scale scenes. Existing representations based on explicit surface suffer from discretization resolution or UV distortion, while implicit volumetric representations lack scalability for large scenes due to the dispersed weight distribution and surface ambiguity. In light of the above challenges, we introduce hash featurized manifold, a novel hash-based featurization coupled with a deferred neural rendering framework. This approach fully unlocks the expressivity of the representation by explicitly concentrating the hash entries on the 2D manifold, thus effectively representing highly detailed contents independent of the discretization resolution. We also introduce a novel dataset, namely GigaNVS, to benchmark cross-scale, high-resolution novel view synthesis of real- world large-scale scenes. Our method significantly outperforms competing baselines on various real-world scenes, yielding an average LPIPS that is ∼ 40% lower than prior state-of-the-art on the challenging GigaNVS benchmark. Please see our project page at: xscalenvs.github.io.

Category-level 6D object pose estimation aims to estimate the rotation, translation and size of unseen instances within specific categories. In this area, dense correspondence-based methods have achieved leading performance. However, they do not explicitly consider the local and global geometric information of different instances, resulting in poor generalization ability to unseen instances with significant shape variations.To deal with this problem, we propose a novel Instance-$\textbf{A}$daptive and $\textbf{G}$eometric-Aware Keypoint Learning method for category-level 6D object pose estimation (AG-Pose), which includes two key designs: (1) The first design is an Instance-Adaptive Keypoint Detection module, which can adaptively detect a set of sparse keypoints for various instances to represent their geometric structures. (2) The second design is a Geometric-Aware Feature Aggregation module, which can efficiently integrate the local and global geometric information into keypoint features.These two modules can work together to establish robust keypoint-level correspondences for unseen instances, thus enhancing the generalization ability of the model.Experimental results on CAMERA25 and REAL275 datasets show that the proposed AG-Pose outperforms state-of-the-art methods by a large margin without category-specific shape priors.

In this work, we present RepKPU, an efficient network for point cloud upsampling. We propose to promote upsampling performance by exploiting better shape representation and point generation strategy. Inspired by KPConv, we propose a novel representation called RepKPoints to effectively characterize the local geometry, whose advantages over prior representations are as follows: (1) density-sensitive; (2) large receptive fields; (3) position-adaptive, which makes RepKPoints a generalized form of previous representations. Moreover, we propose a novel paradigm, namely Kernel-to-Displacement generation, for point generation, where point cloud upsampling is reformulated as the deformation of kernel points. Specifically, we propose KP-Queries, which is a set of kernel points with predefined positions and learned features, to serve as the initial state of upsampling. Using cross-attention mechanisms, we achieve interactions between RepKPoints and KP-Queries, and subsequently KP-Queries are converted to displacement features, followed by a MLP to predict the new positions of KP-Queries which serve as the generated points. Extensive experimental results demonstrate that RepKPU outperforms state-of-the-art methods on several widely-used benchmark datasets with high efficiency.

Point cloud registration is still a challenging and open problem. For example, when the overlap between two point clouds is extremely low, geo-only features may be not sufficient. Therefore, it is important to further explore how to utilize color data in this task. Under such circumstances, we propose ColorPCR for color point cloud registration with multi-stage geometric-color fusion. We design a Hierarchical Color Enhanced Feature Extraction module to extract multi-level geometric-color features, and a GeoColor Superpoint Matching Module to encode transformation-invariant geo-color global context for robust patch correspondences. In this way, both geometric and color data can be used, thus leading to robust performance even under extremely challenging scenarios, such as low overlap between two point clouds. To evaluate the performance of our method, we colorize 3DMatch/3DLoMatch datasets as Color3DMatch/Color3DLoMatch and evaluations on these datasets demonstrate the effectiveness of our proposed method. Our method achieves state-of-the-art registration recall of $\textbf{97.5}$%/$\textbf{88.9}$% on them.

This paper proposes ConsistDreamer - a novel framework that lifts 2D diffusion models with 3D awareness and 3D consistency, thus enabling high-fidelity instruction-guided scene editing. To overcome the fundamental limitation of missing 3D consistency in 2D diffusion models, our key insight is to introduce three synergetic strategies that augment the input of the 2D diffusion model to become 3D-aware and to explicitly enforce 3D consistency during the training process. Specifically, we design surrounding views as context-rich input for the 2D diffusion model, and generate 3D-consistent, structured noise instead of image-independent noise. Moreover, we introduce self-supervised consistency-enforcing training within the per-scene editing procedure. Extensive evaluation shows that our ConsistDreamer achieves state-of-the-art performance for instruction-guided scene editing across various scenes and editing instructions, particularly in complicated large-scale indoor scenes from ScanNet++, with significantly improved sharpness and fine-grained textures. Notably, ConsistDreamer stands as the first work capable of successfully editing complex (e.g., plaid/checkered) patterns. Our project page is at immortalco.github.io/ConsistDreamer.

We propose SceneTex, a novel method for effectively generating high-quality and style-consistent textures for indoor scenes using depth-to-image diffusion priors. Unlike previous methods that either iteratively warp 2D views onto a mesh surface or distillate diffusion latent features without accurate geometric and style cues, SceneTex formulates the texture synthesis task as an optimization problem in the RGB space where style and geometry consistency are properly reflected. At its core, SceneTex proposes a multiresolution texture field to implicitly encode the mesh appearance. We optimize the target texture via a score-distillation-based objective function in respective RGB renderings. To further secure the style consistency across views, we introduce a cross-attention decoder to predict the RGB values by cross-attending to the pre-sampled reference locations in each instance. SceneTex enables various and accurate texture synthesis for 3D-FRONT scenes, demonstrating significant improvements in visual quality and prompt fidelity over the prior texture generation methods.

We present a neural radiance field method for urban-scale semantic and building-level instance segmentation from aerial images by lifting noisy 2D labels to 3D. This is a challenging problem due to two primary reasons. Firstly, objects in urban aerial images exhibit substantial variations in size, including buildings, cars, and roads, which pose a significant challenge for accurate 2D segmentation. Secondly, the 2D labels generated by existing segmentation methods suffer from the multi-view inconsistency problem, especially in the case of aerial images, where each image captures only a small portion of the entire scene. To overcome these limitations, we first introduce a scale-adaptive semantic label fusion strategy that enhances the segmentation of objects of varying sizes by combining labels predicted from different altitudes, harnessing the novel-view synthesis capabilities of NeRF. We then introduce a novel cross-view instance label grouping based on the 3D scene representation to mitigate the multi-view inconsistency problem in the 2D instance labels. Furthermore, we exploit multi-view reconstructed depth priors to improve the geometric quality of the reconstructed radiance field, resulting in enhanced segmentation results. Experiments on multiple real-world urban-scale datasets demonstrate that our approach outperforms existing methods, highlighting its effectiveness. Our code will be made publicly available.

The encoder-decoder network (ED-Net) is a commonly employed choice for existing depth completion methods, but its working mechanism is ambiguous.In this paper, we visualize the internal feature maps to analyze how the network densifies the input sparse depth.We find that the encoder feature of ED-Net focus on the areas with input depth points around.To obtain a dense feature and thus estimate complete depth, the decoder feature tends to complement and enhance the encoder feature by skip-connection to make the fused encoder-decoder feature dense, resulting in the decoder feature also exhibits sparse.However, ED-Net obtains the sparse decoder feature from the dense fused feature at the previous stage, where the ``dense$\Rightarrow$sparse'' process destroys the completeness of features and loses information.To address this issue, we present a depth feature upsampling network (DFU) that explicitly utilizes these dense features to guide the upsampling of a low-resolution (LR) depth feature to a high-resolution (HR) one.The completeness of features is maintained throughout the upsampling process, thus avoiding information loss.Furthermore, we propose a confidence-aware guidance module (CGM), which is confidence-aware and performs guidance with adaptive receptive fields (GARF), to fully exploit the potential of these dense features as guidance.Experimental results show that our DFU, a plug-and-play module, can significantly improve the performance of existing ED-Net based methods with limited computational overheads, and new SOTA results are achieved.Besides, the generalization capability on sparser depth is also enhanced.Project page: https://npucvr.github.io/DFU.

We present ZeroRF, a novel per-scene optimization method addressing the challenge of sparse view 360° reconstruction in neural field representations. Current breakthroughs like Neural Radiance Fields (NeRF) have demonstrated high-fidelity image synthesis but struggle with sparse input views. Existing methods, such as Generalizable NeRFs and per-scene optimization approaches, face limitations in data dependency, computational cost, and generalization across diverse scenarios.To overcome these challenges, we propose ZeroRF, whose key idea is to integrate a tailored Deep Image Prior into a factorized NeRF representation. Unlike traditional methods, ZeroRF parametrizes feature grids with a neural network generator, enabling efficient sparse view 360° reconstruction without any pretraining or additional regularization. Extensive experiments showcase ZeroRF's versatility and superiority in terms of both quality and speed, achieving state-of-the-art results on benchmark datasets.ZeroRF's significance extends to applications in 3D content generation and editing.We will release the code after the paper is published.

We estimate the radiance field of large-scale dynamic areas from multiple vehicle captures under varying environmental conditions. Previous works in this domain are either restricted to static environments, do not scale to more than a single short video, or struggle to separately represent dynamic object instances. To this end, we present a novel, decomposable radiance field approach for dynamic urban environments. We propose a multi-level neural scene graph representation that scales to thousands of images from dozens of sequences with hundreds of fast-moving objects. To enable efficient training and rendering of our representation, we develop a fast composite ray sampling and rendering scheme. To test our approach in urban driving scenarios, we introduce a new, novel view synthesis benchmark. We show that our approach outperforms prior art by a significant margin on both established and our proposed benchmark while being faster in training and rendering.

We introduce Gaussian-Flow, a novel point-based approach for fast dynamic scene reconstruction and real-time rendering from both multi-view and monocular videos. In contrast to the prevalent NeRF-based approaches hampered by slow training and rendering speeds, our approach harnesses recent advancements in point-based 3D Gaussian Splatting (3DGS). Specifically, a novel Dual-Domain Deformation Model (DDDM) is proposed to explicitly model attribute deformations of each Gaussian point, where the time-dependent residual of each attribute is captured by a polynomial fitting in the time domain, and a Fourier series fitting in the frequency domain. The proposed DDDM is capable of modeling complex scene deformations across long video footage, eliminating the need for training separate 3DGS for each frame or introducing an additional implicit neural field to model 3D dynamics. Moreover, the explicit deformation modeling for discretized Gaussian points ensures ultra-fast training and rendering of a 4D scene, which is comparable to the original 3DGS designed for static 3D reconstruction. Our proposed approach showcases a substantial efficiency improvement, achieving a $5\times$ faster training speed compared to the per-frame 3DGS modeling. In addition, quantitative results demonstrate that the proposed Gaussian-Flow significantly outperforms previous leading methods in novel view rendering quality.

3D visual language multi-modal modeling plays an important role in actual human-computer interaction. However, the inaccessibility of large scale 3D-language pairs restricts their applicability in real-world scenarios. In this paper, we aim to handle a real-time multi-task for 6-DoF pose tracking of unknown objects, leveraging 3D-language pre-training scheme from a series of 3D point cloud video streams, while simultaneously performing 3D shape reconstruction in current observation. To this end, we present a generic $\underline{L}$anguage-to-$\underline{4D}$ modeling paradigm termed L4D-Track, that tackles zero-shot 6-DoF $\underline{Track}$ing and shape reconstruction by learning pairwise implicit 3D representation and multi-level multi-modal alignment. Our method constitutes two core parts. 1) Pairwise Implicit 3D Space Representation, that establishes spatial-temporal to language coherence descriptions across continuous 3D point cloud video. 2) Language-to-4D Association and Contrastive Alignment, enables multi-modality semantic connections between 3D point cloud video and language. Our method trained exclusively on public NOCS-REAL275 dataset, achieves promising results on both two publicly benchmarks. This not only shows powerful generalization performance, but also proves its remarkable capability in zero-shot inference. The project is released at https://s-jingtao.github.io/L4D-Track/.

Novel-view synthesis of specular objects like shiny metals or glossy paints remains a significant challenge.Not only the glossy appearance but also global illumination effects, including reflections of other objects in the environment, are critical components to faithfully reproduce a scene.In this paper, we present Neural Directional Encoding (NDE), a view-dependent appearance encoding of neural radiance fields (NeRF) for rendering specular objects.NDE transfers the concept of feature-grid-based spatial encoding to the angular domain, significantly improving the ability to model high-frequency angular signals.In contrast to previous methods that use encoding functions with only angular input, we additionally cone-trace spatial features to obtain a spatially varying directional encoding, which addresses the challenging interreflection effects.Extensive experiments on both synthetic and real datasets show that a NeRF model with NDE (1) outperforms the state of the art on view synthesis of specular objects, and (2) works with small networks to allow fast (real-time) inference.The source code is available at: https://github.com/lwwu2/nde

We propose SNI-SLAM, a semantic SLAM system utilizing neural implicit representation, that simultaneously performs accurate semantic mapping, high-quality surface reconstruction, and robust camera tracking. In this system, we introduce hierarchical semantic representation to allow multi-level semantic comprehension for top-down structured semantic mapping of the scene. In addition, to fully utilize the correlation between multiple attributes of the environment, we integrate appearance, geometry and semantic features through cross-attention for feature collaboration. This strategy enables a more multifaceted understanding of the environment, thereby allowing SNI-SLAM to remain robust even when single attribute is defective. Then, we design an internal fusion-based decoder to obtain semantic, RGB, Truncated Signed Distance Field (TSDF) values from multi-level features for accurate decoding. Furthermore, we propose a feature loss to update the scene representation at the feature level. Compared with low-level losses such as RGB loss and depth loss, our feature loss is capable of guiding the network optimization on a higher-level. Our SNI-SLAM method demonstrates superior performance over all recent NeRF-based SLAM methods in terms of mapping and tracking accuracy on Replica and ScanNet datasets, while also showing excellent capabilities in accurate semantic segmentation and real-time semantic mapping. Codes will be available at https://github.com/IRMVLab/SNI-SLAM.

Multi-camera-based 3D object detection has made notable progress in the past several years. However, we observe that there are cases (e.g. faraway regions) in which popular 2D object detectors are more reliable than state-of-the-art 3D detectors. In this paper, to improve the performance of query-based 3D object detectors, we present a novel query generating approach termed QAF2D, which infers 3D query anchors from 2D detection results. A 2D bounding box of an object in an image is lifted to a set of 3D anchors by associating each sampled point within the box with depth, yaw angle, and size candidates. Then, the validity of each 3D anchor is verified by comparing its projection in the image with its corresponding 2D box, and only valid anchors are kept and used to construct queries. The class information of the 2D bounding box associated with each query is also utilized to match the predicted boxes with ground truth for the set-based loss. The image feature extraction backbone is shared between the 3D detector and 2D detector by adding a small number of prompt parameters. We integrate QAF2D into three popular query-based 3D object detectors and carry out comprehensive evaluations on the nuScenes dataset. The largest improvement that QAF2D can bring about on the nuScenes validation subset is 2.3% NDS and 2.7% mAP. Code is available at https://github.com/nullmax-vision/QAF2D.

Neural radiance fields have achieved remarkable performance in modeling the appearance of 3D scenes. However, existing approaches still struggle with the view-dependent appearance of glossy surfaces, especially under complex lighting of indoor environments. Unlike existing methods, which typically assume distant lighting like an environment map, we propose a learnable Gaussian directional encoding to better model the view-dependent effects under near-field lighting conditions. Importantly, our new directional encoding captures the spatially-varying nature of near-field lighting and emulates the behavior of prefiltered environment maps. As a result, it enables the efficient evaluation of preconvolved specular color at any 3D location with varying roughness coefficients. We further introduce a data-driven geometry prior that helps alleviate the shape radiance ambiguity in reflection modeling. We show that our Gaussian directional encoding and geometry prior significantly improve the modeling of challenging specular reflections in neural radiance fields, which helps decompose appearance into more physically meaningful components.

This paper presents a novel non-rigid point set registration method that is inspired by unsupervised clustering analysis. Unlike previous approaches that treat the source and target point sets as separate entities, we develop a holistic framework where they are formulated as clustering centroids and clustering members, separately. We then adopt Tikhonov regularization with an $\ell_1$-induced Laplacian kernel instead of the commonly used Gaussian kernel to ensure smooth and more robust displacement fields. Our formulation delivers closed-form solutions, theoretical guarantees, independence from dimensions, and the ability to handle large deformations. Subsequently, we introduce a clustering-improved Nystr{\"o}m method to effectively reduce the computational complexity and storage of the Gram matrix to linear, while providing a rigorous bound for the low-rank approximation. Our method achieves high accuracy results across various scenarios and surpasses competitors by a significant margin, particularly on shapes with substantial deformations. Additionally, we demonstrate the versatility of our method in challenging tasks such as shape transfer and medical registration.

Recent years have witnessed the remarkable progress of 3D multi-modality object detection methods based on the Bird's-Eye-View (BEV) perspective. However, most of them overlook the complementary interaction and guidance between LiDAR and camera. In this work, we propose a novel multi-modality 3D objection detection method, named GAFusion, with LiDAR-guided global interaction and adaptive fusion. Specifically, we introduce sparse depth guidance (SDG) and LiDAR occupancy guidance (LOG) to generate 3D features with sufficient depth information. In the following, LiDAR-guided adaptive fusion transformer (LGAFT) is developed to adaptively enhance the interaction of different modal BEV features from a global perspective. Meanwhile, additional downsampling with sparse height compression and multi-scale dual-path transformer (MSDPT) are designed to enlarge the receptive fields of different modal features. Finally, a temporal fusion module is introduced to aggregate features from previous frames. GAFusion achieves state-of-the-art 3D object detection results with 73.6% mAP and 74.9% NDS on the nuScenes test set.

Real-world objects and environments are predominantly composed of edge features, including straight lines and curves. Such edges are crucial elements for various applications, such as CAD modeling, surface meshing, lane mapping, etc. However, existing traditional methods only prioritize lines over curves for simplicity in geometric modeling. To this end, we introduce EMAP, a new method for learning 3D edge representations with a focus on both lines and curves. Our method implicitly encodes 3D edge distance and direction in Unsigned Distance Functions (UDF) from multi-view edge maps. On top of this neural representation, we propose an edge extraction algorithm that robustly abstracts parametric 3D edges from the inferred edge points and their directions. Comprehensive evaluations demonstrate that our method achieves better 3D edge reconstruction on multiple challenging datasets. We further show that our learned UDF field enhances neural surface reconstruction by capturing more details.

Neural implicit fields have been a de facto standard in novel view synthesis. Recently, there exist some methods exploring fusing multiple modalities within a single field, aiming to share implicit features from different modalities to enhance reconstruction performance. However, these modalities often exhibit misaligned behaviors: optimizing for one modality, such as LiDAR, can adversely affect another, like camera performance, and vice versa. In this work, we conduct comprehensive analyses on the multimodal implicit field of LiDAR-camera joint synthesis, revealing the underlying issue lies in the misalignment of different sensors.Furthermore, we introduce AlignMiF, a geometrically aligned multimodal implicit field with two proposed modules: Geometry-Aware Alignment (GAA) and Shared Geometry Initialization (SGI). These modules effectively align the coarse geometry across different modalities, significantly enhancing the fusion process between LiDAR and camera data. Through extensive experiments across various datasets and scenes, we demonstrate the effectiveness of our approach in facilitating better interaction between LiDAR and camera modalities within a unified neural field. Specifically, our proposed AlignMiF, achieves remarkable improvement over recent implicit fusion methods (+2.01 and +3.11 image PSNR on the KITTI-360 and Waymo datasets) and consistently surpasses single modality performance (13.8\% and 14.2\% reduction in LiDAR Chamfer Distance on the respective datasets).

Lidar has become a cornerstone sensing modality for 3D vision, especially for large outdoor scenarios and autonomous driving. Conventional lidar sensors are capable of providing centimeter-accurate distance information by emitting laser pulses into a scene and measuring the time-of-flight (ToF) of the reflection. However, the polarization of the received light that depends on the surface orientation and material properties is usually not considered. As such, the polarization modality has the potential to improve scene reconstruction beyond distance measurements. In this work, we introduce a novel long-range polarization wavefront lidar sensor (PolLidar) that modulates the polarization of the emitted and received light. Departing from conventional lidar sensors, PolLidar allows access to the raw time-resolved polarimetric wavefronts. We leverage polarimetric wavefronts to estimate normals, distance, and material properties in outdoor scenarios with a novel learned reconstruction method. To train and evaluate the method, we introduce a simulated and real-world long-range dataset with paired raw lidar data, ground truth distance, and normal maps. We find that the proposed method improves normal and distance reconstruction by 53% mean angular error and 41% mean absolute error compared to existing shape-from-polarization (SfP) and ToF methods. Code and data are open-sourced here.

Estimating full-body human motion via sparse tracking signals from head-mounted displays and hand controllers in 3D scenes is crucial to applications in AR/VR. One of the biggest challenges to this task is the one-to-many mapping from sparse observations to dense full-body motions, which endowed inherent ambiguities. To help resolve this ambiguous problem, we introduce a new framework to combine rich contextual information provided by scenes to benefit full-body motion tracking from sparse observations. To estimate plausible human motions given sparse tracking signals and 3D scenes, we develop $\text{S}^2$Fusion, a unified framework fusing scene and sparse signals with a conditional diffusion model. $\text{S}^2$Fusion first extracts the spatial-temporal relations resided in the sparse signals via a periodic autoencoder, and then produces time-alignment feature embedding as additional inputs. Subsequently, by drawing initial noisy motion from a pre-trained prior, $\text{S}^2$Fusion utilizes conditional diffusion to fuse scene geometry and sparse tracking signals to generate full-body scene-aware motions. The sampling procedure of $\text{S}^2$Fusion is further guided by a specially designed scene-penetration loss and phase-matching loss, which effectively regularizes the motion of the lower body even in the absence of any tracking signals, making generated motion much more plausible and coherent. Extensive experimental results have demonstrated that our $\text{S}^2$Fusion outperforms the state-of-the-art in terms of estimation quality and smoothness.

We introduce FaceTalk, a novel generative approach designed for synthesizing high-fidelity 3D motion sequences of talking human heads from input audio signal. To capture the expressive, detailed nature of human heads, including hair, ears, and finer-scale eye movements, we propose to couple speech signal with the latent space of neural parametric head models to create high-fidelity, temporally coherent motion sequences. We propose a new latent diffusion model for this task, operating in the expression space of neural parametric head models, to synthesize audio-driven realistic head sequences. In the absence of a dataset with corresponding NPHM expressions to audio, we optimize for these correspondences to produce a dataset of temporally-optimized NPHM expressions fit to audio-video recordings of people talking. To the best of our knowledge, this is the first work to propose a generative approach for realistic and high-quality motion synthesis of volumetric human heads, representing a significant advancement in the field of audio-driven 3D animation. Notably, our approach stands out in its ability to generate plausible motion sequences that can produce high-fidelity head animation coupled with the NPHM shape space. Our experimental results substantiate the effectiveness of FaceTalk, consistently achieving superior and visually natural motion, encompassing diverse facial expressions and styles, outperforming existing methods by 75% in perceptual user study evaluation

The emergence of Neural Radiance Fields (NeRF) has greatly impacted 3D scene modeling and novel-view synthesis. As a kind of visual media for 3D scene representation, compression with high rate-distortion performance is an eternal target. Motivated by advances in neural compression and neural field representation, we propose NeRFCodec, an end-to-end NeRF compression framework that integrates non-linear transform, quantization, and entropy coding for memory-efficient scene representation. Since training a non-linear transform directly on a large scale of NeRF feature planes is impractical, we discover that pre-trained neural 2D image codec can be utilized for compressing the features when adding content-specific parameters. Specifically, we reuse neural 2D image codec but modify its encoder and decoder heads, while keeping the other parts of the pre-trained decoder frozen. This allows us to train the full pipeline via supervision of rendering loss and entropy loss, yielding the rate-distortion balance by updating the content-specific parameters. At test time, the bitstreams containing latent code, feature decoder head, and other side information are transmitted for communication. Experimental results demonstrate our method outperforms existing NeRF compression methods, enabling high-quality novel view synthesis with a memory budget of 0.5 MB.

In dynamic 3D environments, the ability to recognize a diverse range of objects without the constraints of predefined categories is indispensable for real-world applications. In response to this need, we introduce OV3D, an innovative framework designed for open-vocabulary 3D semantic segmentation. OV3D leverages the broad open-world knowledge embedded in vision and language foundation models to establish a fine-grained correspondence between 3D points and textual entity descriptions. These entity descriptions are enriched with contextual information, enabling a more open and comprehensive understanding. By seamlessly aligning 3D point features with entity text features, OV3D empowers open-vocabulary recognition in the 3D domain, achieving state-of-the-art open-vocabulary semantic segmentation performance across multiple datasets, including ScanNet, Matterport3D, and nuScenes. Code will be available.

As pretrained text-to-image diffusion models become increasingly powerful, recent efforts have been made to distill knowledge from these text-to-image pretrained models for optimizing a text-guided 3D model. Most of the existing methods generate a holistic 3D model from a plain text input. This can be problematic when the text describes a complex scene with multiple objects, because the vectorized text embeddings are inherently unable to capture a complex description with multiple entities and relationships. Holistic 3D modeling of the entire scene further prevents accurate grounding of text entities and concepts. To address this limitation, we propose GraphDreamer, a novel framework to generate compositional 3D scenes from scene graphs, where objects are represented as nodes and their interactions as edges. By exploiting node and edge information in scene graphs, our method makes better use of the pretrained text-to-image diffusion model and is able to fully disentangle different objects without image-level supervision. To facilitate modeling of object-wise relationships, we use signed distance fields as representation and impose a constraint to avoid inter-penetration of objects. To avoid manual scene graph creation, we design a text prompt for ChatGPT to generate scene graphs based on text inputs. We conduct both qualitative and quantitative experiments to validate the effectiveness of GraphDreamer in generating high-fidelity compositional 3D scenes with disentangled object entities.

The booming of 3D recognition in the 2020s began with the introduction of point cloud transformers. They quickly overwhelmed sparse CNNs and became state-of-the-art models, especially in 3D semantic segmentation. However, sparse CNNs are still valuable networks, due to their efficiency treasure, and ease of application. In this work, we reexamine the design distinctions and test the limits of what a sparse CNN can achieve. We discover that the key credit to the performance difference is adaptivity. Specifically, we propose two key components, i.e., adaptive receptive fields (spatially) and adaptive relation, to bridge the gap. This exploration led to the creation of Omni-Adaptive 3D CNNs (OA-CNNs), a family of networks that integrates a lightweight module to greatly enhance the adaptivity of sparse CNNs at minimal computational cost. Without any self-attention modules, OA-CNNs favorably surpass point transformers in terms of accuracy in both indoor and outdoor scenes, with much less latency and memory cost. Notably, it achieves 76.1%, 78.9%, and 70.6% mIoU on ScanNet v2, nuScenes, and SemanticKITTI validation benchmarks respectively, while maintaining at most 5x better speed than transformer counterparts. This revelation highlights the potential of pure sparse CNNs to outperform transformer-related networks. Our code is built upon Pointcept, which is available at https://github.com/Pointcept/Pointcept.

We revisit certain problems of pose estimation based on 3D--2D correspondences between features which may be points or lines. Specifically, we address the two previously-studied minimal problems of estimating camera extrinsics from $p \in \{ 1, 2 \}$ point--point correspondences and $l=3-p$ line--line correspondences. To the best of our knowledge, all of the previously-known practical solutions to these problems required computing the roots of degree $\ge 4$ (univariate) polynomials when $p=2$, or degree $\ge 8$ polynomials when $p=1.$ We describe and implement two elementary solutions which reduce the degrees of the needed polynomials from $4$ to $2$ and from $8$ to $4$, respectively. We show experimentally that the resulting solvers are numerically stable and fast: when compared to the previous state-of-the art, we obtain nearly an order of magnitude speedup.

This paper introduces CN-RMA, a novel approach for 3D indoor object detection from multi-view images. We observe the key challenge as the ambiguity of image and 3D correspondence without explicit geometry to provide occlusion information. To address this issue, CN-RMA leverages the synergy of 3D reconstruction networks and 3D object detection networks, where the reconstruction network provides a rough Truncated Signed Distance Function (TSDF) and guides image features to vote to 3D space correctly in an end-to-end manner. Specifically, we associate weights to sampled points of each ray through ray marching, representing the contribution of a pixel in an image to corresponding 3D locations. Such weights are determined by the predicted signed distances so that image features vote only to regions near the reconstructed surface. Our method achieves state-of-the-art performance in 3D object detection from multi-view images, as measured by mAP@0.25 and mAP@0.5 on the ScanNet and ARKitScenes datasets. The code and models are released at https://github.com/SerCharles/CN-RMA.

Holistic understanding of urban scenes based on RGB images is a challenging yet important problem. It encompasses understanding both the geometry and appearance to enable novel view synthesis, parsing semantic labels, and tracking moving objects. Despite considerable progress, existing approaches often focus on specific aspects of this task and require additional inputs such as LiDAR scans or manually annotated 3D bounding boxes. In this paper, we introduce a novel pipeline that utilizes 3D Gaussian Splatting for holistic urban scene understanding. Our main idea involves the joint optimization of geometry, appearance, semantics, and motion using a combination of static and dynamic 3D Gaussians, where moving object poses are regularized via physical constraints. Our approach offers the ability to render new viewpoints in real-time, yielding 2D and 3D semantic information with high accuracy, and reconstruct dynamic scenes, even in scenarios where 3D bounding box detection are highly noisy. Experimental results on KITTI, KITTI-360, and Virtual KITTI 2 demonstrate the effectiveness of our approach.

Implicit neural representation (INR), in combination with geometric rendering, has recently been employed in real-time dense RGB-D SLAM.Despite active research endeavors being made, there lacks a unified protocol for fair evaluation, impeding theevolution of this area. In this work, we establish, to our knowledge, the first open-source benchmark framework toevaluate the performance of a wide spectrum of commonly used INRs and rendering functions for mapping and localization. The goal of our benchmark is to 1) gain an intuition of how different INRs and rendering functions impact mapping and localization and 2) establish a unified evaluation protocol w.r.t. the design choices that may impact the mapping and localization. With the framework, we conduct a large suite of experiments, offering various insights in choosing the INRs and geometricrendering functions: for example, the dense feature grid outperforms other INRs (e.g. tri-plane and hash grid), even when geometric and color features are jointly encoded for memory efficiency. To extend the findings into the practical scenario, a hybrid encoding strategy is proposed to bring the best of the accuracy and completion from the grid-based and decomposition-based INRs. We further propose explicit hybrid encoding for high-fidelity dense grid mapping to comply with the RGB-D SLAM system that puts the premise on robustness and computation efficiency.

Dense simultaneous localization and mapping (SLAM) is crucial for robotics and augmented reality applications. However, current methods are often hampered by the non-volumetric or implicit way they represent a scene. This work introduces SplaTAM, an approach that, for the first time, leverages explicit volumetric representations, i.e., 3D Gaussians, to enable high-fidelity reconstruction from a single unposed RGB-D camera, surpassing the capabilities of existing methods. SplaTAM employs a simple online tracking and mapping system tailored to the underlying Gaussian representation. It utilizes a silhouette mask to elegantly capture the presence of scene density. This combination enables several benefits over prior representations, including fast rendering and dense optimization, quickly determining if areas have been previously mapped, and structured map expansion by adding more Gaussians. Extensive experiments show that SplaTAM achieves up to 2x superior performance in camera pose estimation, map construction, and novel-view synthesis over existing methods, paving the way for more immersive high-fidelity SLAM applications.

In recent years, there has been an explosion of 2D vision models for numerous tasks such as semantic segmentation, style transfer or scene editing, enabled by large-scale 2D image datasets. At the same time, there has been renewed interest in 3D scene representations such as neural radiance fields from multi-view images. However, the availability of 3D or multiview data is still substantially limited compared to 2D image datasets, making extending 2D vision model to 3D data highly desirable but also very challenging. Indeed, extending a single 2D vision operator like scene editing to 3D typically requires a highly creative method specialized to that task and often requires per-scene optimization. In this paper, we ask the question of whether any 2D vision model can be lifted to make 3D consistent predictions. We answer this question in the affirmative; our new Lift3D method trains to predict unseen views on feature spaces generated by a few visual models (i.e. DINO and CLIP), but then generalizes to novel vision operators and tasks, such as style transfer, super-resolution, open vocabulary segmentation and image colorization; for some of these tasks, there is no comparable previous 3D method. In many cases, we even outperform state-of-the-art methods specialized for the task in question. Moreover, Lift3D is a zero-shot method, in the sense that it requires no task-specific training, nor scene-specific optimization.

This work proposes a novel representation of injective deformations of 3D space, which overcomes existing limitations of injective methods, namely inaccuracy, lack of robustness, and incompatibility with general learning and optimization frameworks. Our core idea is to reduce the problem to a ``deep'' composition of multiple 2D mesh-based piecewise-linear maps. Namely, we build differentiable layers that produce mesh deformations through Tutte's embedding (guaranteed to be injective in 2D), and compose these layers over different planes to create complex 3D injective deformations of the 3D volume. We show that our method provides the ability to efficiently and accurately optimize and learn complex deformations, outperforming other injective approaches. As a main application, we show our ability to produce complex and artifact-free NeRF deformations.

Since its proposal, Neural Radiance Fields (NeRF) has achieved great success in related tasks, mainly adopting the hierarchical volume sampling (HVS) strategy for volume rendering. However, the HVS of NeRF approximates distributions using piecewise constant functions, which provides a relatively rough estimation. Based on the observation that a well-trained weight function $w(t)$ and the $L_0$ distance between points and the surface have very high similarity, we propose $L_0$-Sampler by incorporating the $L_0$ model into $w(t)$ to guide the sampling process. Specifically, we propose using piecewise exponential functions rather than piecewise constant functions for interpolation, which can not only approximate quasi-$L_0$ weight distributions along rays quite well but can be easily implemented with a few lines of code change without additional computational burden. Stable performance improvements can be achieved by applying $L_0$-Sampler to NeRF and related tasks like 3D reconstruction. Code is available at \href{https://ustc3dv.github.io/L0-Sampler/}{https://ustc3dv.github.io/L0-Sampler/}.

Automatic text-to-3D generation that combines Score Distillation Sampling (SDS) with the optimization of volume rendering has achieved remarkable progress in synthesizing realistic 3D objects. Yet most existing text-to-3D methods by SDS and volume rendering suffer from inaccurate geometry, e.g., the Janus issue, since it is hard to explicitly integrate 3D priors into implicit 3D representations. Besides, it is usually time-consuming for them to generate elaborate 3D models with rich colors. In response, this paper proposes GSGEN, a novel method that adopts Gaussian Splatting, a recent state-of-the-art representation, to text-to-3D generation. GSGEN aims at generating high-quality 3D objects and addressing existing shortcomings by exploiting the explicit nature of Gaussian Splatting that enables the incorporation of 3D prior. Specifically, our method adopts a progressive optimization strategy, which includes a geometry optimization stage and an appearance refinement stage. In geometry optimization, a coarse representation is established under 3D point cloud diffusion prior along with the ordinary 2D SDS optimization, ensuring a sensible and 3D-consistent rough shape. Subsequently, the obtained Gaussians undergo an iterative appearance refinement to enrich texture details. In this stage, we increase the number of Gaussians by compactness-based densification to enhance continuity and improve fidelity. With these designs, our approach can generate 3D assets with delicate details and accurate geometry. Extensive evaluations demonstrate the effectiveness of our method, especially for capturing high-frequency components.

The limited scale of current 3D shape datasets hinders the advancements in 3D shape understanding, and motivates multi-modal learning approaches which transfer learned knowledge from data-abundant 2D image and language modalities to 3D shapes. However, even though the image and language representations have been aligned by cross-modal models like CLIP, we find that the image modality fails to contribute as much as the language in existing multi-modal 3D representation learning methods. This is attributed to the domain shift in the 2D images and the distinct focus of each modality. To more effectively leverage both modalities in the pre-training, we introduce TriAdapter Multi-Modal Learning (TAMM) – a novel two-stage learning approach based on three synergetic adapters. First, our CLIP Image Adapter mitigates the domain gap between 3D-rendered images and natural images, by adapting the visual representations of CLIP for synthetic image-text pairs. Subsequently, our Dual Adapters decouple the 3D shape representation space into two complementary sub-spaces: one focusing on visual attributes and the other for semantic understanding, which ensure a more comprehensive and effective multi-modal pre-training. Extensive experiments demonstrate that TAMM consistently enhances 3D representations for a wide range of 3D encoder architectures, pre-training datasets, and downstream tasks. Notably, we boost the zero-shot classification accuracy on Objaverse-LVIS from 46.8 to 50.7, and improve the 5-way 10-shot linear probing classification accuracy on ModelNet40 from 96.1 to 99.0. Project page: https://alanzhangcs.github.io/tamm-page.

3D Gaussian splatting has achieved very impressive performance in real-time novel view synthesis. However, it often suffers from over-reconstruction during Gaussian densification where high-variance image regions are covered by a few large Gaussians only, leading to blur and artifacts in the rendered images. We design a progressive frequency regularization (FreGS) technique to tackle the over-reconstruction issue within the frequency space. Specifically, FreGS performs coarse-to-fine Gaussian densification by exploiting low-to-high frequency components that can be easily extracted with low-pass and high-pass filters in the Fourier space. By minimizing the discrepancy between the frequency spectrum of the rendered image and the corresponding ground truth, it achieves high-quality Gaussian densification and alleviates the over-reconstruction of Gaussian splatting effectively. Experiments over multiple widely adopted benchmarks (e.g., Mip-NeRF360, Tanks-and-Temples and Deep Blending) show that FreGS achieves superior novel view synthesis and outperforms the state-of-the-art consistently.

Multi-view inverse rendering is the problem of estimating the scene parameters such as shapes, materials, or illuminations from a sequence of images captured under different viewpoints. Many approaches, however, assume single light bounce and thus fail to recover challenging scenarios like inter-reflections. On the other hand, simply extending those methods to consider multi-bounced light requires more assumptions to alleviate the ambiguity. To address this problem, we propose Neural Incident Stokes Fields (NeISF), a multi-view inverse rendering framework that reduces ambiguities using polarization cues. The primary motivation for using polarization cues is that it is the accumulation of multi-bounced light, providing rich information about geometry and material. Based on this knowledge, the proposed incident Stokes field efficiently models the accumulated polarization effect with the aid of an original physically-based differentiable polarimetric renderer. Lastly, experimental results show that our method outperforms the existing works in synthetic and real scenarios.

Even though Non-rigid Structure-from-Motion (NRSfM) has been extensively studied and great progress has been made, there are still key challenges that hinder their broad real-world applications: 1) the inherent motion/rotation ambiguity requires either explicit camera motion recovery with extra constraint or complex Procrustean Alignment; 2) existing low-rank modeling of the global shape can over-penalize drastic deformations in the 3D shape sequence. This paper proposes to resolve the above issues from a spatial-temporal modeling perspective. First, we propose a novel Temporally-smooth Procrustean Alignment module that estimates 3D deforming shapes and adjusts the camera motion by aligning the 3D shape sequence consecutively. Our new alignment module remedies the requirement of complex reference 3D shape during alignment, which is more conductive to non-isotropic deformation modeling. Second, we propose a spatial-weighted approach to enforce the low-rank constraint adaptively at different locations to accommodate drastic spatially-variant deformation reconstruction better. Our modeling outperform existing low-rank based methods, and extensive experiments across different datasets validate the effectiveness of our method.

A neural signed distance function (SDF) is a convenient shape representation for many tasks, such as surface reconstruction, editing and generation.However, neural SDFs are difficult to fit to raw point clouds, such as those sampled from the surface of a shape by a scanner.A major issue occurs when the shape's geometry is very different from the structural biases implicit in the network's initialization.In this case, we observe that the standard loss formulation does not guide the network towards the correct SDF values.We circumvent this problem by introducing guiding points, and use them to steer the optimization towards the true shape via small incremental changes for which the loss formulation has a good descent direction.We show that this point-guided homotopy-based optimization scheme facilitates a deformation from an easy problem to the difficult reconstruction problem.We also propose a metric to quantify the difference in surface geometry between a target shape and an initial surface, which helps indicate whether the standard loss formulation is guiding towards the target shape.Our method outperforms previous state-of-the-art approaches, with large improvements on shapes identified by this metric as particularly challenging.

Neural Radiance Fields (NeRF) have shown impressive capabilities for photorealistic novel view synthesis when trained on dense inputs. However, when trained on sparse inputs, NeRF typically encounters issues of incorrect density or color predictions, mainly due to insufficient coverage of the scene causing partial and sparse supervision, thus leading to significant performance degradation. While existing works mainly consider ray-level consistency to construct 2D learning regularization based on rendered color, depth, or semantics on image planes, in this paper we propose a novel approach that models 3D spatial radiance consistency to improve NeRF's performance with sparse inputs. Specifically, we first adopt a voxel-based ray sampling strategy to ensure that the sampled rays intersect with a certain voxel in 3D space. We then randomly sample additional points within the voxel and apply a Transformer to infer the properties of other points on each ray, which are then incorporated into the volume rendering. By backpropagating through the rendering loss, we enhance the consistency among neighboring points. Additionally, we propose to use a contrastive loss on the encoder output of the Transformer to further improve consistency within each voxel. Experiments demonstrate that our method yields significant improvement over different radiance fields in the sparse inputs setting, and achieves comparable performance with current works. Our models and codes will be made publicly available upon acceptance.

3D editing plays a crucial role in many areas such as gaming and virtual reality. Traditional 3D editing methods, which rely on representations like meshes and point clouds, often fall short in realistically depicting complex scenes. On the other hand, methods based on implicit 3D representations, like Neural Radiance Field (NeRF), render complex scenes effectively but suffer from slow processing speeds and limited control over specific scene areas.In response to these challenges, our paper presents GaussianEditor, an innovative and efficient 3D editing algorithm based on Gaussian Splatting (GS), a novel 3D representation.GaussianEditor enhances precision and control in editing through our proposed Gaussian semantic tracing, which traces the editing target throughout the training process.Additionally, we propose Hierarchical Gaussian splatting (HGS) to achieve stabilized and fine results under stochastic generative guidance from 2D diffusion models.We also develop editing strategies for efficient object removal and integration, a challenging task for existing methods.Our comprehensive experiments demonstrate GaussianEditor's superior control, efficacy, and rapid performance, marking a significant advancement in 3D editing.

Understanding how the surrounding environment changes is crucial for performing downstream tasks safely and reliably in autonomous driving applications. Recent occupancy estimation techniques using only camera images as input can provide dense occupancy representations of large-scale scenes based on the current observation. However, they are mostly limited to representing the current 3D space and do not consider the future state of surrounding objects along the time axis. To extend camera-only occupancy estimation into spatiotemporal prediction, we propose Cam4DOcc, a new benchmark for camera-only 4D occupancy forecasting, evaluating the surrounding scene changes in a near future. We build our benchmark based on multiple publicly available datasets, including nuScenes, nuScenes-Occupancy, and Lyft-Level5, which provides sequential occupancy states of general movable and static objects, as well as their 3D backward centripetal flow. To establish this benchmark for future research with comprehensive comparisons, we introduce four baseline types from diverse camera-based perception and prediction implementations, including a static-world occupancy model, voxelization of point cloud prediction, 2D-3D instance-based prediction, and our proposed novel end-to-end 4D occupancy forecasting network. Furthermore, the standardized evaluation protocol for preset multiple tasks is also provided to compare the performance of all the proposed baselines on present and future occupancy estimation with respect to objects of interest in autonomous driving scenarios. The dataset and our implementation of all four baselines in the proposed Cam4DOcc benchmark will be released as open source.

Diffusion models have shown remarkable results for image generation, editing and inpainting. Recent works explore diffusion models for 3D shape generation with neural implicit functions, i.e., signed distance function and occupancy function. However, they are limited to shapes with closed surfaces, which prevents them from generating diverse 3D real-world contents containing open surfaces. In this work, we present UDiFF, a 3D diffusion model for unsigned distance fields (UDFs) which is capable to generate textured 3D shapes with open surfaces from text conditions or unconditionally. Our key idea is to generate UDFs in spatial-frequency domain with an optimal wavelet transformation, which produces a compact representation space for UDF generation. Specifically, instead of searching for an appropriate wavelet transformation which requires expensive manual efforts and still leads to large information loss, we propose a data-driven approach to learn the optimal wavelet transformation for UDFs. We evaluate UDiFF to show our advantages by numerical and visual comparisons with the latest methods on widely used benchmarks.

We introduce the Panoptic 3D Reconstruction task, a unified and holistic scene understanding task for a monocular video. And we present PanoRecon - a novel framework to address this new task, which realizes an online geometry reconstruction alone with dense semantic and instance labeling. Specifically, PanoRecon incrementally performs panoptic 3D reconstruction for each video fragment consisting of multiple consecutive key frames, from a volumetric feature representation using feed-forward neural networks. We adopt a depth-guided back-projection strategy to sparse and purify the volumetric feature representation. We further introduce a voxel clustering module to get object instances in each local fragment, and then design a tracking and fusion algorithm for the integration of instances from different fragments to ensure temporal coherence. Such design enables our PanoRecon to yield a coherent and accurate panoptic 3D reconstruction. Experiments on ScanNetV2 demonstrate a very competitive geometry reconstruction result compared with state-of-the-art reconstruction methods, as well as promising 3D panoptic segmentation result with only RGB input, while being real-time. Code will be publicly available upon acceptance.

Self-supervised image backbones can be used to address complex 2D tasks (e.g., semantic segmentation, object discovery) very efficiently and with little or no downstream supervision. Ideally, 3D backbones for lidar should be able to inherit these properties after distillation of these powerful 2D features. The most recent methods for image-to-lidar distillation on autonomous driving data show promising results, obtained thanks to distillation methods that keep improving. Yet, we still notice a large performance gap when measuring the quality of distilled and fully supervised features by linear probing. In this work, instead of focusing only on the distillation method, we study the effect of three pillars for distillation: the 3D backbone, the pretrained 2D backbones, and the pretraining dataset. In particular, thanks to our scalable distillation method named ScaLR, we show that scaling the 2D and 3D backbones and pretraining on diverse datasets leads to a substantial improvement of the feature quality. This allows us to significantly reduce the gap between the quality of distilled and fully-supervised 3D features, and to improve the robustness of the pretrained backbones to domain gaps and perturbations.

Grouping is inherently ambiguous due to the multiple levels of granularity in which one can decompose a scene --- should the wheels of an excavator be considered separate or part of the whole? We present Group Anything with Radiance Fields (GARField), an approach for decomposing 3D scenes into a hierarchy of semantically meaningful groups from posed image inputs. To do this we embrace group ambiguity through physical scale: by optimizing a scale-conditioned 3D affinity feature field, a point in the world can belong to different groups of different sizes. We optimize this field from a set of 2D masks provided by Segment Anything (SAM) in a way that respects coarse-to-fine hierarchy, using scale to consistently fuse conflicting masks from different viewpoints. From this field we can derive a hierarchy of possible groupings via automatic tree construction or user interaction. We evaluate GARField on a variety of in-the-wild scenes and find it effectively extracts groups at many levels: clusters of objects, objects, and various subparts. GARField inherently represents multi-view consistent groupings and produces higher fidelity groups than the input SAM masks. GARField's hierarchical grouping could have exciting downstream applications such as 3D asset extraction or dynamic scene understanding.

While recent depth completion methods have achieved remarkable results filling in relatively dense depth maps (e.g., projected 64-line LiDAR on KITTI or 500 sampled points on NYUv2) with RGB guidance, their performance on very sparse input (e.g., 4-line LiDAR or 32 depth point measurements) is unverified. These sparser regimes present new challenges, as a 4-line LiDAR increases the distance between pixels without depth and their nearest depth point sixfold from 5 pixels to 30 pixels compared to 64 lines. Observing that existing methods struggle with sparse and variable distribution depth maps, we propose an Affinity-Based Shift Correction (ASC) module that iteratively aligns depth predictions to input depth based on predicted affinities between image pixels and depth points. Our framework enables each depth point to adaptively influence and improve predictions across the image, leading to largely improved results for fewer-line, fewer-point, and variable sparsity settings. Further, we show improved performance in domain transfer from KITTI to nuScenes and from random sampling to irregular point distributions. Our correction module can easily be added to any depth completion or RGB-only depth estimation model, notably allowing the latter to perform both completion and estimation with a single model.

3D reconstruction methods such as Neural Radiance Fields (NeRFs) excel at rendering photorealistic novel views of complex scenes. However, recovering a high-quality NeRF typically requires tens to hundreds of input images, resulting in a time-consuming capture process. We present ReconFusion to reconstruct real-world scenes using only a few photos. Our approach leverages a diffusion prior for novel view synthesis, trained on synthetic and multiview datasets, which regularizes a NeRF-based 3D reconstruction pipeline at novel camera poses beyond those captured by the set of input images. Our method synthesizes realistic geometry and texture in underconstrained regions while preserving the appearance of observed regions. We perform an extensive evaluation across various real-world datasets, including forward-facing and 360-degree scenes, demonstrating significant performance improvements over previous few-view NeRF reconstruction approaches.

Scene coordinate regression (SCR) methods are a family of visual localization methods that directly regress 2D-3D matches for camera pose estimation. They are effective in small-scale scenes but face significant challenges in large-scale scenes that are further amplified in the absence of ground truth 3D point clouds for supervision. Here, the model can only rely on reprojection constraints and needs to implicitly triangulate the points. The challenges stem from a fundamental dilemma: The network has to be invariant to observations of the same landmark at different viewpoints and lighting conditions, etc., but at the same time discriminate unrelated but similar observations. The latter becomes more relevant and severe in larger scenes. In this work, we tackle this problem by introducing the concept of co-visibility to the network. We propose GLACE, which integrates pre-trained global and local encodings and enables SCR to scale to large scenes with only a single small-sized network. Specifically, we propose a novel feature diffusion technique that implicitly groups the reprojection constraints with co-visibility and avoids overfitting to trivial solutions. Additionally, our position decoder parameterizes the output positions for large-scale scenes more effectively. Without using 3D models or depth maps for supervision, our method achieves state-of-the-art results on large-scale scenes with a low-map-size model. On Cambridge landmarks, we achieve a 17% lower median position error than Poker, the ensemble of state-of-the-art SCR method ACE, with a single model.

We present NARUTO, a neural active reconstruction system that combines a hybrid neural representation with uncertainty learning, enabling high-fidelity surface reconstruction. Our approach leverages a multi-resolution hash-grid as the mapping backbone, chosen for its exceptional convergence speed and capacity to capture high-frequency local features.The centerpiece of our work is the incorporation of an uncertainty learning module that dynamically quantifies reconstruction uncertainty while actively reconstructing the environment. By harnessing learned uncertainty, we propose a novel uncertainty aggregation strategy for goal searching and efficient path planning. Our system autonomously explores by targeting uncertain observations and reconstructs environments with remarkable completeness and fidelity. We also demonstrate the utility of this uncertainty-aware approach by enhancing SOTA neural SLAM systems through an active ray sampling strategy.Extensive evaluations of NARUTO in various environments, using an indoor scene simulator, confirm its superior performance and state-of-the-art status in active reconstruction, as evidenced by its impressive results on benchmark datasets like Replica and MP3D.

The integration of neural rendering and the SLAM system recently showed promising results in joint localization and photorealistic view reconstruction. However, existing methods, fully relying on implicit representations, are so resource-hungry that they cannot run on portable devices, which deviates from the original intention of SLAM. In this paper, we present Photo-SLAM, a novel SLAM framework with a hyper primitives map. Specifically, we simultaneously exploit explicit geometric features for localization and learn implicit photometric features to represent the texture information of the observed environment. In addition to actively densifying hyper primitives based on geometric features, we further introduce a Gaussian-Pyramid-based training method to progressively learn multi-level features, enhancing photorealistic mapping performance. The extensive experiments with monocular, stereo, and RGB-D datasets prove that our proposed system Photo-SLAM significantly outperforms current state-of-the-art SLAM systems for online photorealistic mapping, e.g., PSNR is 30\% higher and rendering speed is hundreds of times faster in the Replica dataset. Moreover, the Photo-SLAM can run at real-time speed using an embedded platform such as Jetson AGX Orin, showing the potential of robotics applications. Project Page and code: https://huajianup.github.io/research/Photo-SLAM/.

We propose a new structure-from-motion framework to recover accurate camera poses and point clouds from unordered images. Traditional SfM systems typically rely on the successful detection of repeatable keypoints across multiple views as the first step, which is difficult for texture-poor scenes, and poor keypoint detection may break down the whole SfM system. We propose a new detector-free SfM framework to draw benefits from the recent success of detector-free matchers to avoid the early determination of keypoints, while solving the multiview inconsistency issue of detector-free matchers.Specifically, our framework first reconstructs a coarse SfM model from quantized detector-free matches. Then, it refines the model by a novel iterative refinement pipeline, which iterates between an attention-based multiview matching module to refine feature tracks and a geometry refinement module to improve the reconstruction accuracy. Experiments demonstrate that the proposed framework outperforms existing detector-based SfM systems on common benchmark datasets. We also collect a texture-poor SfM dataset to demonstrate the capability of our framework to reconstruct texture-poor scenes. Our code and dataset will be released for reproducibility.

In this paper, we propose a new framework for online 3D scene perception. Conventional 3D scene perception methods are offline, i.e., take an already reconstructed 3D scene geometry as input, which is not applicable in robotic applications where the input data is streaming RGB-D videos rather than a complete 3D scene reconstructed from pre-collected RGB-D videos.To deal with online 3D scene perception tasks where data collection and perception should be performed simultaneously, the model should be able to process 3D scenes frame by frame and make use of the temporal information.To this end, we propose an adapter-based plug-and-play module for the backbone of 3D scene perception model, which constructs memory to cache and aggregate the extracted RGB-D features to empower offline models with temporal learning ability. Specifically, we propose a queued memory mechanism to cache the supporting point cloud and image features. Then we devise aggregation modules which directly perform on the memory and pass temporal information to current frame. We further propose 3D-to-2D adapter to enhance image features with strong global context. Our adapters can be easily inserted into mainstream offline architectures of different tasks and significantly boost their performance on online tasks. Extensive experiments on ScanNet and SceneNN datasets demonstrate our approach achieves leading performance on three 3D scene perception tasks compared with state-of-th-art online methods by simply finetuning existing offline models, without any model and task-specific designs.

Vision-centric 3D environment understanding is both vital and challenging for autonomous driving systems. Recently, object-free methods have attracted considerable attention. Such methods perceive the world by predicting the semantics of discrete voxel grids but fail to construct continuous and accurate obstacle surfaces. To this end, in this paper, we propose SurroundSDF to implicitly predict the signed distance field (SDF) and semantic field for the continuous perception from surround images. Specifically, we introduce a query-based approach and utilize SDF constrained by the Eikonal formulation to accurately describe the surfaces of obstacles. Furthermore, considering the absence of precise SDF ground truth, we propose a novel weakly supervised paradigm for SDF, referred to as the Sandwich Eikonal formulation, which emphasizes applying correct and dense constraints on both sides of the surface, thereby enhancing the perceptual accuracy of the surface. Experiments suggest that our method achieves SOTA for both occupancy prediction and 3D scene reconstruction tasks on the nuScenes dataset. The code will be released when paper is accepted.

Capturing and re-animating the 3D structure of articulated objects present significant barriers. On one hand, methods requiring extensively calibrated multi-view setups are prohibitively complex and resource-intensive, limiting their practical applicability. On the other hand, while single-camera Neural Radiance Fields (NeRFs) offer a more streamlined approach, they have excessive training and rendering costs. 3D Gaussian Splatting would be a suitable alternative, but for two reasons. Firstly, 3D dynamic Gaussians require synchronized multi-view cameras, and secondly, the lack of controllability in dynamic scenarios. We present CoGS, a method for Controllable Gaussian Splatting, that enables the direct manipulation of scene elements, offering real-time control of dynamic scenes without the prerequisite of pre-computing control signals. We evaluated CoGS using both synthetic and real-world datasets that include dynamic objects that differ in degree of difficulty.In our evaluations, CoGS consistently outperformed existing dynamic and controllable neural representations in terms of visual fidelity.

We present DrivingGaussian, an efficient and effective framework for surrounding dynamic autonomous driving scenes. For complex scenes with moving objects, we first sequentially and progressively model the static background of the entire scene with incremental static 3D Gaussians. We then leverage a composite dynamic Gaussian graph to handle multiple moving objects, individually reconstructing each object and restoring their accurate positions and occlusion relationships within the scene. We further use a LiDAR prior for Gaussian Splatting to reconstruct scenes with greater details and maintain panoramic consistency. DrivingGaussian outperforms existing methods in driving scene reconstruction and enables photorealistic surround-view synthesis with high-fidelity and multi-camera consistency. The source code and trained models will be released.

We propose GS-IR, a novel inverse rendering approach based on 3D Gaussian Splatting (GS) that leverages forward mapping volume rendering to achieve photorealistic novel view synthesis and relighting results. Unlike previous works that use implicit neural representations and volume rendering (e.g. NeRF), which suffer from low expressive power and high computational complexity, we extend GS, a top-performance representation for novel view synthesis, to estimate scene geometry, surface material, and environment illumination from multi-view images captured under unknown lighting conditions. There are two main problems when introducing GS to inverse rendering: 1) GS does not support producing plausible normal natively; 2) forward mapping (e.g. rasterization and splatting) cannot trace the occlusion like backward mapping (e.g. ray tracing). To address these challenges, our GS-IR proposes an efficient optimization scheme that incorporates a depth-derivation-based regularization for normal estimation and a baking-based occlusion to model indirect lighting. The flexible and expressive GS representation allows us to achieve fast and compact geometry reconstruction, photorealistic novel view synthesis, and effective physically-based rendering. We demonstrate the superiority of our method over baseline methods through qualitative and quantitative evaluations on various challenging scenes.

Gated cameras flood-illuminate a scene and capture the time-gated impulse response of a scene. By employing nanosecond-scale gates, existing sensors are capable of capturing mega-pixel gated images, delivering dense depth improving on today's LiDAR sensors in spatial resolution and depth precision. Although gated depth estimation methods deliver a million of depth estimates per frame, their resolution is still an order below existing RGB imaging methods. In this work, we combine high-resolution stereo HDR RCCB cameras with gated imaging, allowing us to exploit depth cues from active gating, multi-view RGB and multi-view NIR sensing -- multi-view and gated cues across the entire spectrum. The resulting capture system consists only of low-cost CMOS sensors and flood-illumination. We propose a novel stereo-depth estimation method that is capable of exploiting these multi-modal multi-view depth cues, including the active illumination that is measured by the RCCB camera when removing the IR-cut filter. The proposed method achieves accurate depth at long ranges up to 220 m, outperforming the next best existing method by 16\% in MAE on accumulated LiDAR ground-truth.

We present a novel method for efficiently producing semi-dense matches across images.Previous detector-free matcher LoFTR has shown remarkable matching capability in handling large-viewpoint change and texture-poor scenarios but suffers from low efficiency.We revisit its design choices and derive multiple improvements for both efficiency and accuracy.One key observation is that performing the transformer over the entire feature map is redundant due to shared local information, therefore we propose an aggregated attention mechanism with adaptive token selection for efficiency.Furthermore, we find spatial bias exists in LoFTR's fine correlation module, which is adverse to matching accuracy.A novel two-stage correlation layer is proposed to achieve unbiased subpixel correspondences for accuracy improvement.Our efficiency optimized model is $\sim 2.5\times$ faster than LoFTR which can even surpass state-of-the-art efficient sparse matching pipeline SuperPoint + LightGlue. Moreover, extensive experiments show that our method can achieve higher accuracy compared with competitive semi-dense matchers, with considerable efficiency benefits.This opens up exciting prospects for large-scale or latency-sensitive applications such as image retrieval and 3D reconstruction.Our code will be released for reproducibility.

3D scene representations have gained immense popularity in recent years. Methods that use Neural Radiance fields are versatile for traditional tasks such as novel view synthesis. In recent times, some work has emerged that aims to extend the functionality of NeRF beyond view synthesis, for semantically aware tasks such as editing and segmentation using 3D feature field distillation from 2D foundation models. However, these methods have two major limitations: (a) they are limited by the rendering speed of NeRF pipelines, and (b) implicitly represented feature fields suffer from continuity artifacts reducing feature quality. Recently, 3D Gaussian Splatting has shown state-of-the-art performance on real-time radiance field rendering. In this work, we go one step further: in addition to radiance field rendering, we enable 3D Gaussian splatting on arbitrary-dimension semantic features via 2D foundation model distillation. This translation is not straightforward: naively incorporating feature fields in the 3DGS framework encounters significant challenges, notably the disparities in spatial resolution and channel consistency between RGB images and feature maps. We propose architectural and training changes to efficiently avert this problem. Our proposed method is general, and our experiments showcase novel view semantic segmentation, language-guided editing and segment anything through learning feature fields from state-of-the-art 2D foundation models such as SAM and CLIP-LSeg. Across experiments, our distillation method is able to provide comparable or better results, while being significantly faster to both train and render. Additionally, to the best of our knowledge, we are the first method to enable point and bounding-box prompting for radiance field manipulation, by leveraging the SAM model. Project website at: https://feature-3dgs.github.io/

Structure-from-motion (SfM) is a long-standing problem in the computer vision community, which aims to reconstruct the camera poses and 3D structure of a scene from a set of unconstrained 2D images. Classical frameworks solve this problem in an incremental manner by detecting and matching keypoints, registering images, triangulating 3D points, and conducting bundle adjustment. Recent research efforts have predominantly revolved around harnessing the power of deep learning techniques to enhance specific elements (e.g., keypoint matching), but are still based on the original, non-differentiable pipeline. Instead, we propose a new deep SfM pipeline, where each component is fully differentiable and thus can be trained in an end-to-end manner. To this end, we introduce new mechanisms and simplifications. First, we build on recent advances in deep 2D point tracking to extract reliable pixel-accurate tracks, which eliminates the need for chaining pairwise matches. Furthermore, we recover all cameras simultaneously based on the image and track features instead of gradually registering cameras. Finally, we optimise the cameras and triangulate 3D points via a differentiable bundle adjustment layer. We attain state-of-the-art performance on three popular datasets, CO3D, IMC Phototourism, and ETH3D.

Point Cloud Registration is a critical and challenging task in computer vision. Recent advancements have predominantly embraced a coarse-to-fine matching mechanism, with the key to matching the superpoints located in patches with inter-frame consistent structures. However, previous methods still face challenges with ambiguous matching, because the interference information aggregated from irrelevant regions may disturb the capture of inter-frame consistency relations, leading to wrong matches. To address this issue, we propose Dynamic Cues-Assisted Transformer (DCATr). Firstly, the interference from irrelevant regions is greatly reduced by constraining attention to certain cues, i.e., regions with highly correlated structures of potential corresponding superpoints. Secondly, cues-assisted attention is designed to mine the inter-frame consistency relations, while more attention is assigned to pairs with high consistent confidence in feature aggregation. Finally, a dynamic updating fashion is proposed to facilitate mining richer consistency information, further improving aggregated features' distinctiveness and relieving matching ambiguity. Extensive evaluations on indoor and outdoor standard benchmarks demonstrate that DCATr outperforms all state-of-the-art methods.

This study addresses the challenge of performing visual localization in demanding conditions such as night-time scenarios, adverse weather, and seasonal changes. While many prior studies have focused on improving image matching performance to facilitate reliable dense keypoint matching between images, existing methods often heavily rely on predefined feature points on a reconstructed 3D model. Consequently, they tend to overlook unobserved keypoints during the matching process. Therefore, dense keypoint matches are not fully exploited, leading to a notable reduction in accuracy, particularly in noisy scenes. To tackle this issue, we propose a novel localization method that extracts reliable semi-dense 2D-3D matching points based on dense keypoint matches. This approach involves regressing semi-dense 2D keypoints into 3D scene coordinates using a point inference network. The network utilizes both geometric and visual cues to effectively infer 3D coordinates for unobserved keypoints from the observed ones. The abundance of matching information significantly enhances the accuracy of camera pose estimation, even in scenarios involving noisy or sparse 3D models. Comprehensive evaluations demonstrate that the proposed method outperforms other methods in challenging scenes and achieves competitive results in large-scale visual localization benchmarks. The code will be available at https://github.com/TruongKhang/DeViLoc.

Applying NeRF to downstream perception tasks for scene understanding and representation is becoming increasingly popular. Most existing methods treat semantic prediction as an additional rendering task, \textit{i.e.}, the "label rendering" task, to build semantic NeRFs.However, by rendering semantic/instance labels per pixel without considering the contextual information of the rendered image, these methods usually suffer from unclear boundary segmentation and abnormal segmentation of pixels within an object. To solve this problem, we propose Generalized Perception NeRF (GP-NeRF), a novel pipeline that makes the widely used segmentation model and NeRF work compatibly under a unified framework, for facilitating context-aware 3D scene perception. To accomplish this goal, we introduce Transformers to aggregate radiance as well as semantic embedding fields jointly for novel views and facilitate the joint volumetric rendering upon both fields.In addition, we propose two self-distillation mechanisms, i.e., the Semantic Distill Loss and the Depth-Guided Semantic Distill Loss, to enhance the discrimination and quality of the semantic field and maintenance of geometric consistency.In evaluation, we conduct experimental comparisons under two perception tasks (\textit{i.e.} semantic and instance segmentation) using both synthetic and real-world datasets. Notably, our method outperforms SOTA approaches by 6.94\%, 11.76\%, and 8.47\% on generalized semantic segmentation, finetuning semantic segmentation, and instance segmentation, respectively.

Neural Radiance Fields (NeRFs) have demonstrated remarkable potential in capturing complex 3D scenes with high fidelity. However, one persistent challenge that hinders the widespread adoption of NeRFs is the computational bottleneck due to the volumetric rendering. On the other hand, 3D Gaussian splatting (3DGS) has recently emerged as an alternative representation that leverages a 3D Gaussisan-based representation and adopts the rasterization pipeline to render the images rather than volumetric rendering, achieving very fast rendering speed and promising image quality. However, a significant drawback arises as 3DGS entails a substantial number of 3D Gaussians to maintain the high fidelity of the rendered images, which requires a large amount of memory and storage. To address this critical issue, we place a specific emphasis on two key objectives: reducing the number of Gaussian points without sacrificing performance and compressing the Gaussian attributes, such as view-dependent color and covariance. To this end, we propose a learnable mask strategy that significantly reduces the number of Gaussians while preserving high performance. In addition, we propose a compact but effective representation of view-dependent color by employing a grid-based neural field rather than relying on spherical harmonics. Finally, we learn codebooks to compactly represent the geometric attributes of Gaussian by vector quantization. With model compression techniques such as quantization and entropy coding, we consistently show over 25$\times$ reduced storage and enhanced rendering speed, while maintaining the quality of the scene representation, compared to 3DGS. Our work provides a comprehensive framework for 3D scene representation, achieving high performance, fast training, compactness, and real-time rendering. Our project page is available at https://maincold2.github.io/c3dgs/.

Implicit Neural Representations have gained prominence as a powerful framework for capturing complex data modalities, encompassing a wide range from 3D shapes to images and audio. Within the realm of 3D shape representation, Neural Signed Distance Functions (SDF) have demonstrated remarkable potential in faithfully encoding intricate shape geometry. However, learning SDFs from 3D point clouds in the absence of ground truth supervision remains a very challenging task. In this paper, we propose a method to infer occupancy fields instead of SDFs as they are easier to learn from sparse inputs. We leverage a margin-based uncertainty measure to differentiably sampling from the decision boundary of the occupancy function and supervise the sampled boundary points using the input point cloud. We further stabilise the optimization process at the early stages of the training by biasing the occupancy function towards minimal entropy fields while maximizing its entropy at the input point cloud. Through extensive experiments and evaluations, we illustrate the efficacy of our proposed method, highlighting its capacity to improve implicit shape inference with respect to baselines and the state-of-the-art using synthetic and real data.

Many contemporary studies utilize grid-based models for neural field representation, but a systematic analysis of grid-based models is still missing, hindering the improvement of those models. Therefore, this paper introduces a theoretical framework for grid-based models. This framework points out that these models' approximation and generalization behaviors are determined by grid tangent kernels (GTK), which are intrinsic properties of grid-based models. The proposed framework facilitates a consistent and systematic analysis of diverse grid-based models. Furthermore, the introduced framework motivates the development of a novel grid-based model named the Multiplicative Fourier Adaptive Grid (MulFAGrid). The numerical analysis demonstrates that MulFAGrid exhibits a lower generalization bound than its predecessors, indicating its robust generalization performance. Empirical studies reveal that MulFAGrid achieves state-of-the-art performance in various tasks, including 2D image fitting, 3D signed distance field (SDF) reconstruction, and novel view synthesis, demonstrating superior representation ability. The project website is available at~\href{https://sites.google.com/view/cvpr24-2034-submission/home}{this link}.

Humans commonly work with multiple objects in daily life and can intuitively transfer manipulation skills to novel objects by understanding object functional regularities. However, existing technical approaches for analyzing and synthesizing hand-object manipulation are mostly limited to handling a single hand and object due to the lack of data support. To address this, we construct TACO, an extensive bimanual hand-object-interaction dataset spanning a large variety of tool-action-object compositions for daily human activities. TACO contains 2.5K motion sequences paired with third-person and egocentric views, precise hand-object 3D meshes, and action labels. To rapidly expand the data scale, we present a fully automatic data acquisition pipeline combining multi-view sensing with an optical motion capture system. With the vast research fields provided by TACO, we benchmark three generalizable hand-object-interaction tasks: compositional action recognition, generalizable hand-object motion forecasting, and cooperative grasp synthesis. Extensive experiments reveal new insights, challenges, and opportunities for advancing the studies of generalizable hand-object motion analysis and synthesis. Our data and code are available at https://taco2024.github.io.

We establish rigorous benchmarks for visual perception robustness. Synthetic images such as ImageNet-C, ImageNet-9, and Stylized ImageNet provide specific type of evaluation over synthetic corruptions, backgrounds, and textures, yet those robustness benchmarks are restricted in specified variations and have low synthetic quality. In this work, we introduce generative model as a data source for synthesizing hard images that benchmark deep models' robustness. Leveraging diffusion models, we are able to generate images with more diversified backgrounds, textures, and materials than any prior work, where we term this benchmark as ImageNet-D. Experimental results show that ImageNet-D results in a significant accuracy drop to a range of vision models, from the standard ResNet visual classifier to the latest foundation models like CLIP and MiniGPT-4, significantly reducing their accuracy by up to 60\%. Our work suggests that diffusion models can be an effective source to test vision models. The code and dataset are available at https://github.com/chenshuang-zhang/imagenet_d.

To advance research in learning-based defogging algorithms, various synthetic fog datasets have been developed. However, exsiting datasets created using the Atmospheric Scattering Model (ASM) or real-time rendering engines often struggle to produce photo-realistic foggy images that accurately mimic the actual imaging process. This limitation hinders the effective generalization of models from synthetic to real data. In this paper, we introduce an end-to-end simulation pipeline designed to generate photo-realistic foggy images. This pipeline comprehensively considers the entire physically-based foggy scene imaging process, closely aligning with real-world image capture methods. Based on this pipeline, we present a new synthetic fog dataset named SynFog, which features both skylight and active lighting conditions, as well as three levels of fog density. Experimental results demonstrate that models trained on SynFog exhibit superior performance in visual perception and detection accuracy compared to others when applied to real-world foggy images.

Fine-grained action analysis in multi-person sports is complex due to athletes' quick movements and intense physical confrontations, which result in severe visual obstructions in most scenes. In addition, accessible multi-person sports video datasets lack fine-grained action annotations in both space and time, adding to the difficulty in fine-grained action analysis.To this end, we construct a new multi-person basketball sports video dataset named FineSports, which contains fine-grained semantic and spatial-temporal annotations on 10,000 NBA game videos, covering 52 fine-grained action types, 16,000 action instances, and 123,000 spatial-temporal bounding boxes. We also propose a new prompt-driven spatial-temporal action location approach called PoSTAL, composed of a prompt-driven target action encoder (PTA) and an action tube-specific detector (ATD) to directly generate target action tubes with fine-grained action types without any off-line proposal generation. Extensive experiments on the FineSports dataset demonstrate that PoSTAL outperforms state-of-the-art methods. Data and code are available at https://github.com/PKU-ICST-MIPL/FineSports_CVPR2024.

We introduce Infinigen Indoors, a Blender-based procedural generator of photorealistic indoor scenes. It builds upon the existing Infinigen system, which focuses on natural scenes, but expands its coverage to indoor scenes by introducing a diverse library of procedural indoor assets, including furniture, architecture elements, appliances, and other day-to-day objects. It also introduces a constraint-based arrangement system, which consists of a domain-specific language for expressing diverse constraints on scene composition, and a solver that generates scene compositions that maximally satisfy the constraints. Another contribution is an export tool that allows the generated 3D objects and scenes to be directly used for training embodied agents in real-time simulators such as Omniverse and Unreal. Infinigen Indoors will be open-sourced under the BSD license.

Recent advances in large-scale pretraining have yielded visual foundation models with strong generalization abilities. Such models aim to learn representations that are useful for a wide range of downstream tasks such as classification, segmentation, and generation. These models have been proven successful in 2D tasks, where they can delineate and localize objects. But how much do they really understand in 3D? In this work, we analyze the 3D awareness of the representations learned by visual foundation models. We argue that 3D awareness at least implies (1) representing the 3D structure of the visible surface and (2) consistent representations across views. We conduct a series of experiments that analyze the representations learned using task-specific probes and zero-shot inference procedures on frozen features, revealing several limitations of current foundation models.

Video generation has witnessed significant advancements, yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal evaluation system should provide insights to inform future developments of video generation. To this end, we present VBench, a comprehensive benchmark suite that dissects "video generation quality" into specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. VBench has three appealing properties: 1) Comprehensive Dimensions: VBench comprises 16 dimensions in video generation (e.g., subject identity inconsistency, motion smoothness, temporal flickering, and spatial relationship, etc). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception, for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions, and various content types. We also investigate the gaps between video and image generation models. We will open-source VBench, including all prompts, evaluation methods, generated videos, and human preference annotations, and also include more video generation models in VBench to drive forward the field of video generation.

Vision-language generative AI has demonstrated remarkable promise for empowering cross-modal scene understanding of autonomous driving and high-definition (HD) map systems. However, current benchmark datasets lack multi-modal point cloud, image, and language data pairs. Recent approaches utilize visual instruction learning and cross-modal prompt engineering to expand vision-language models into this domain. In this paper, we propose a new vision-language benchmark that can be used to finetune traffic and HD map domain-specific foundation models. Specifically, we annotate and leverage large-scale, broad-coverage traffic and map data extracted from huge HD map annotations, and use CLIP and LLaMA-2 / Vicuna to finetune a baseline model with instruction-following data. Our experimental results across various algorithms reveal that while visual instruction-tuning large language models (LLMs) can effectively learn meaningful representations from MAPLM-QA, there remains significant room for further advancements. To facilitate applying LLMs and multi-modal data into self-driving research, we will release our visual-language QA data, and the baseline models at GitHub.com/LLVM-AD/MAPLM.

The creation of new datasets often presents new challenges for video recognition and can inspire novel ideas while addressing these challenges. While existing datasets mainly comprise landscape mode videos, our paper seeks to introduce portrait mode videos to the research community and highlight the unique challenges associated with this video format. With the growing popularity of smartphones and social media applications, recognizing portrait mode videos is becoming increasingly important. To this end, we have developed the first dataset dedicated to portrait mode video recognition, namely PortraitMode-400. The taxonomy of PortraitMode-400 was constructed in a data-driven manner, comprising 400 fine-grained categories, and rigorous quality assurance was implemented to ensure the accuracy of human annotations. In addition to the new dataset, we conducted a comprehensive analysis of the impact of video format (portrait mode versus landscape mode) on recognition accuracy and spatial bias due to the different formats. Furthermore, we designed extensive experiments to explore key aspects of portrait mode video recognition, including the choice of data augmentation, evaluation procedure, the importance of temporal information, and the role of audio modality. Building on the insights from our experimental results and the introduction of PortraitMode-400, our paper aims to inspire further research efforts in this emerging research area.

Foot contact is an important cue for human motion capture, understanding, and generation. Existing datasets tend to annotate dense foot contact using visual matching with thresholding or incorporating pressure signals. However, these approaches either suffer from low accuracy or are only designed for small-range and slow motion. There is still a lack of a vision-pressure multimodal dataset with large-range and fast human motion, as well as accurate and dense foot-contact annotation. To fill this gap, we propose a Multimodal MoCap Dataset with Vision and Pressure sensors, named MMVP. MMVP provides accurate and dense plantar pressure signals synchronized with RGBD observations, which is especially useful for both plausible shape estimation, robust pose fitting without foot drifting, and accurate global translation tracking. To validate the dataset, we propose an RGBD-P SMPL fitting method and also a monocular-video-based baseline framework, VP-MoCap, for human motion capture. Experiments demonstrate that our RGBD-P SMPL Fitting results significantly outperform pure visual motion capture. Moreover, VP-MoCap outperforms SOTA methods in foot-contact and global translation estimation accuracy. We believe the configuration of the dataset and the baseline frameworks will stimulate the research in this direction and also provide a good reference for MoCap applications in various domains. Project page: https://metaverse-ai-lab-thu.github.io/MMVP-Dataset/.

Counterfactual reasoning, a fundamental aspect of human cognition, involves contemplating alternatives to established facts or past events, significantly enhancing our abilities in planning and decision-making. In light of the advancements in current multi-modal large language models, we explore their effectiveness in counterfactual reasoning. To facilitate this investigation, we introduce a novel dataset, C-VQA, specifically designed to examine the counterfactual reasoning capabilities of modern multi-modal large language models. This dataset is constructed by infusing original questions with counterfactual presuppositions, spanning various types such as numerical and boolean queries. It encompasses a mix of real and synthetic data, representing a wide range of difficulty levels. Our thorough evaluations of contemporary vision-language models using this dataset have revealed substantial performance drops, with some models showing up to a 40\% decrease, highlighting a significant gap between current models and human-like vision reasoning capabilities. We hope our dataset will serve as a vital benchmark for evaluating the counterfactual reasoning capabilities of models.Code and dataset are publicly available at https://bzhao.me/C-VQA/.

In recent decades, the vision community has witnessed remarkable progress in visual recognition, partially owing to advancements in dataset benchmarks. Notably, the established COCO benchmark has propelled the development of modern detection and segmentation systems. However, the COCO segmentation benchmark has seen comparatively slow improvement over the last decade. Originally equipped with coarse polygon annotations for thing instances, it gradually incorporated coarse superpixel annotations for stuff regions, which were subsequently heuristically amalgamated to yield panoptic segmentation annotations. These annotations, executed by different groups of raters, have resulted not only in coarse segmentation masks but also in inconsistencies between segmentation types. In this study, we undertake a comprehensive reevaluation of the COCO segmentation annotations. By enhancing the annotation quality and expanding the dataset to encompass 383K images with more than 5.18M panoptic masks, we introduce COCONut, the COCO Next Universal segmenTation dataset. COCONut harmonizes segmentation annotations across semantic, instance, and panoptic segmentation with meticulously crafted high-quality masks, and establishes a robust benchmark for all segmentation tasks. To our knowledge, COCONut stands as the inaugural large-scale universal segmentation dataset, verified by human raters. We anticipate that the release of COCONut will significantly contribute to the community's ability to assess the progress of novel neural networks.

Traffic scene perception in computer vision is a critically important task to achieve intelligent cities. To date, most existing datasets focus on autonomous driving scenes. We observe that the models trained on those driving datasets often yield unsatisfactory results on traffic monitoring scenes. However, little effort has been put into improving the traffic monitoring scene understanding, mainly due to the lack of specific datasets. To fill this gap, we introduce a specialized traffic monitoring dataset, termed TSP6K, containing images from the traffic monitoring scenario, with high-quality pixel-level and instance-level annotations. The TSP6K dataset captures more crowded traffic scenes with several times more traffic participants than the existing driving scenes. We perform a detailed analysis of the dataset and comprehensively evaluate previous popular scene parsing methods, instance segmentation methods and unsupervised domain adaption methods. Furthermore, considering the vast difference in instance sizes, we propose a detail refining decoder for scene parsing, which recovers the details of different semantic regions in traffic scenes owing to the proposed TSP6K dataset. Experiments show its effectiveness in parsing the traffic monitoring scenes. Code and dataset are available at https://github.com/PengtaoJiang/TSP6K .

We present a new dataset called Real Acoustic Fields (RAF) that captures real acoustic room data from multiple modalities. The dataset includes high-quality and densely captured room impulse response data paired with multi-view images, and precise 6DoF pose tracking data for sound emitters and listeners in the rooms. We used this dataset to evaluate existing methods for novel-view acoustic synthesis and impulse response generation which previously relied on synthetic data. In our evaluation, we thoroughly assessed existing audio and audio-visual models against multiple criteria and proposed settings to enhance their performance on real-world data. We also conducted experiments to investigate the impact of incorporating visual data (i.e., images and depth) into neural acoustic field models. Additionally, we demonstrated the effectiveness of a simple sim2real approach, where a model is pre-trained with simulated data and fine-tuned with sparse real-world data, resulting in significant improvements in the few-shot learning approach. \ourdata is the first dataset to provide densely captured room acoustic data, making it an ideal resource for researchers working on audio and audio-visual neural acoustic field modeling techniques. We will make our dataset publicly available.

Domain generalization aims to solve the challenge of Out-of-Distribution (OOD) generalization by leveraging common knowledge learned from multiple training domains to generalize to unseen test domains. To accurately evaluate the OOD generalization ability, it is required that test data information is unavailable. However, the current domain generalization protocol may still have potential test data information leakage. This paper examines the risks of test data information leakage from two aspects of the current evaluation protocol: supervised pretraining on ImageNet and oracle model selection. We propose modifications to the current protocol that we should employ self-supervised pretraining or train from scratch instead of employing the current supervised pretraining, and we should use multiple test domains. These would result in a more precise evaluation of OOD generalization ability. We also rerun the algorithms with the modified protocol and introduce new leaderboards to encourage future research in domain generalization with a fairer comparison.

Multimodal summarization with multimodal output (MSMO) has emerged as a promising research direction. Nonetheless, numerous limitations exist within existing public MSMO datasets, including insufficient maintenance, data inaccessibility, limited size, and the absence of proper categorization, which pose significant challenges.To address these challenges and provide a comprehensive dataset for this new direction, we have meticulously curated the \textbf{MMSum} dataset. Our new dataset features (1) Human-validated summaries for both video and textual content, providing superior human instruction and labels for multimodal learning.(2) Comprehensively and meticulously arranged categorization, spanning 17 principal categories and 170 subcategories to encapsulate a diverse array of real-world scenarios.(3) Benchmark tests performed on the proposed dataset to assess various tasks and methods, including \textit{video summarization}, \textit{text summarization}, and \textit{multimodal summarization}. To champion accessibility and collaboration, we will release the \textbf{MMSum} dataset and the data collection tool as fully open-source resources, fostering transparency and accelerating future developments.

The study of complex human interactions and group activities has become a focal point in human-centric computer vision. However, progress in related tasks is often hindered by the challenges of obtaining large-scale labeled datasets from real-world scenarios. To address the limitation, we introduce M3Act, a synthetic data generator for multi-view multi-group multi-person human atomic actions and group activities. Powered by Unity Engine, M3Act features multiple semantic groups, highly diverse and photorealistic images, and a comprehensive set of annotations, which facilitates the learning of human-centered tasks across single-person, multi-person, and multi-group conditions. We demonstrate the advantages of M3Act across three core experiments. The results suggest our synthetic dataset can significantly improve the performance of several downstream methods and replace real-world datasets to reduce cost. Notably, M3Act improves the state-of-the-art MOTRv2 on DanceTrack dataset, leading to a hop on the leaderboard from 10th to 2nd place. Moreover, M3Act opens new research for controllable 3D group activity generation. We define multiple metrics and propose a competitive baseline for the novel task. Our code and data are available at our project page: http://cjerry1243.github.io/M3Act.

Egocentric sensors such as AR/VR devices capture human-object interactions and offer the potential to provide task-assistance by recalling 3D locations of objects of interest in the surrounding environment. This capability requires instance tracking in real-world 3D scenes from egocentric videos (IT3DEgo). We explore this problem by first introducing a new benchmark dataset, consisting of RGB and depth videos, per-frame camera pose, and instance-level annotations in both 2D camera and 3D world coordinates. We present an evaluation protocol which evaluates tracking performance in 3D coordinates with two settings for enrolling instances to track: (1) single-view online enrollment where an instance is specified on-the-fly based on the human wearer's interactions. and (2) multi-view pre-enrollment where images of an instance to be tracked are stored in memory ahead of time. To address IT3DEgo, we first re-purpose methods from relevant areas, e.g., single object tracking (SOT) -- running SOT methods to track instances in 2D frames and lifting them to 3D using camera pose and depth. We also present a simple method that leverages pretrained segmentation and detection models to generate proposals from RGB frames and match proposals with enrolled instance images. Our experiments show that our method (with no finetuning) significantly outperforms SOT-based approaches in the egocentric setting. We conclude by arguing that the problem of egocentric instance tracking is made easier by leveraging camera pose and using a 3D allocentric (world) coordinate representation.

In precision agriculture, the detection and recognition of insects play an essential role in the ability of crops to grow healthy and produce a high-quality yield. The current machine vision model requires a large volume of data to achieve high performance. However, there are approximately 5.5 million different insect species in the world. None of the existing insect datasets can cover even a fraction of them due to varying geographic locations and acquisition costs. In this paper, we introduce a novel ``Insect-1M'' dataset, a game-changing resource poised to revolutionize insect-related foundation model training. Covering a vast spectrum of insect species, our dataset, including 1 million images with dense identification labels of taxonomy hierarchy and insect descriptions, offers a panoramic view of entomology, enabling foundation models to comprehend visual and semantic information about insects like never before. Then, to efficiently establish an Insect Foundation Model, we develop a micro-feature self-supervised learning method with a Patch-wise Relevant Attention mechanism capable of discerning the subtle differences among insect images. In addition, we introduce Description Consistency loss to improve micro-feature modeling via insect descriptions. Through our experiments, we illustrate the effectiveness of our proposed approach in insect modeling and achieve State-of-the-Art performance on standard benchmarks of insect-related tasks. Our Insect Foundation Model and Dataset promise to empower the next generation of insect-related vision models, bringing them closer to the ultimate goal of precision agriculture.

Low-resource settings are well-established in natural language processing, where many languages lack sufficient data for deep learning at scale. However, low-resource problems are under-explored in computer vision. In this paper, we address this gap and explore the challenges of low-resource image tasks with vision foundation models. We first collect a benchmark of genuinely low-resource image data, covering historic maps, circuit diagrams, and mechanical drawings. These low-resource settings all share three challenges: data scarcity, fine-grained differences, and the distribution shift from natural images to the specialized domain of interest. While existing foundation models have shown impressive generalizability, we find they cannot transfer well to our low-resource tasks. To begin to tackle the challenges of low-resource vision, we introduce one simple baseline per challenge. Specifically, we i) enlarge the data space by generative models, ii) adopt the best sub-kernels to encode local regions for fine-grained difference discovery and iii) learn attention for specialized domains. Experiments on our three low-resource tasks demonstrate our proposals already provide a better baseline than transfer learning, data augmentation, and fine-grained methods. This highlights the unique characteristics and challenges of low-resource vision for foundation models that warrant further investigation. Project page: https://xiaobai1217.github.io/Low-Resource-Vision/.

Determining the location of an image anywhere on Earth is a complex visual task, which makes it particularly relevant for evaluating computer vision algorithms. Yet, the absence of standard, large-scale, open-access datasets with reliably localizable images has limited its potential. To address this issue, we introduce OpenStreetView-5M, a large-scale, open-access dataset comprising over 5.1 million geo-referenced street view images, covering 225 countries and territories. In contrast to existing benchmarks, we enforce a strict train/test separation, allowing us to evaluate the relevance of learned geographical features beyond mere memorization.To demonstrate the utility of our dataset, we conduct an extensive benchmark of various state-of-the-art image encoders, spatial representations, and training strategies. All associated codes and models can be found at https://github.com/gastruc/osv5m.

Estimating the 3D structure of the human body from natural scenes is a fundamental aspect of visual perception. 3D human pose estimation is a vital step in advancing fields like AIGC and human-robot interaction, serving as a crucial technique for understanding and interacting with human actions in real-world settings.However, the current datasets, often collected under single laboratory conditions using complex motion capture equipment and unvarying backgrounds, are insufficient. The absence of datasets on variable conditions is stalling the progress of this crucial task. To facilitate the development of 3D pose estimation, we present FreeMan, the first large-scale, multi-view dataset collected under the real-world conditions. FreeMan was captured by synchronizing $8$ smartphones across diverse scenarios. It comprises $11M$ frames from $8000$ sequences, viewed from different perspectives. These sequences cover $40$ subjects across $10$ different scenarios, each with varying lighting conditions. We have also established an semi-automated pipeline containing error detection to reduce the workload of manual check and ensure precise annotation.We provide comprehensive evaluation baselines for a range of tasks, underlining the significant challenges posed by FreeMan. Further evaluations of standard indoor/outdoor human sensing datasets reveal that FreeMan offers robust representation transferability in real and complex scenes. FreeMan is now publicly available at https://wangjiongw.github.io/freeman.

In this paper, we present LiDAR-Net, a new real-scanned indoor point cloud dataset, containing nearly $3.6$ billion precisely point-level annotated points, covering an expansive area of 30,000$m^2$. It encompasses three prevalent daily environments, including learning scenes, working scenes, and living scenes. LiDAR-Net is characterized by its non-uniform point distribution, e.g., scanning holes and scanning lines. Additionally, it meticulously records and annotates scanning anomalies, including reflection noise and ghost. These anomalies stem from specular reflections on glass or metal, as well as distortions due to moving persons. LiDAR-Net's realistic representation of non-uniform distribution and anomalies significantly enhances the training of deep learning models, leading to improved generalization in practical applications. We thoroughly evaluate the performance of state-of-the-art algorithms on LiDAR-Net and provide a detailed analysis of the results. Crucially, our research identifies several fundamental challenges in understanding indoor point clouds, contributing essential insights to future explorations in this field. Our dataset can be found online: http://lidar-net.njumeta.com

Existing person re-identification methods have achieved remarkable advances in appearance-based identity association across homogeneous cameras, such as ground-ground matching. However, as a more practical scenario, aerial-ground person re-identification (AGReID) among heterogeneous cameras has received minimal attention. To alleviate the disruption of discriminative identity representation by dramatic view discrepancy as the most significant challenge in AGReID, the view-decoupled transformer (VDT) is proposed as a simple yet effective framework. Two major components are designed in VDT to decouple view-related and view-unrelated features, namely hierarchical subtractive separation and orthogonal loss, where the former separates these two features inside the VDT, and the latter constrains these two to be independent. In addition, we contribute a large-scale AGReID dataset called CARGO, consisting of five/eight aerial/ground cameras, 5,000 identities, and 108,563 images. Experiments on two datasets show that VDT is a feasible and effective solution for AGReID, surpassing the previous method on mAP/Rank1 by up to 5.0\%/2.7\% on CARGO and 3.7\%/5.2\% on AG-ReID, keeping the same magnitude of computational complexity. Our dataset and code will be released after the review process.

Existing text-based person retrieval datasets often have relatively coarse-grained text annotations. This hinders the model to comprehend the fine-grained semantics of query texts in real scenarios. To address this problem, we contribute a new benchmark named UFineBench for text-based person retrieval with ultra-fine granularity.Firstly, we construct a new dataset named UFine6926. We collect a large number of person images and manually annotate each image with two detailed textual descriptions, averaging 80.8 words each. The average word count is three to four times that of the previous datasets. In addition of standard in-domain evaluation, we also propose a special evaluation paradigm more representative of real scenarios. It contains a new evaluation set with cross domains, cross textual granularity and cross textual styles, named UFine3C, and a new evaluation metric for accurately measuring retrieval ability, named mean Similarity Distribution (mSD). Moreover, we propose CFAM, a more efficient algorithm especially designed for text-based person retrieval with ultra fine-grained texts. It achieves fine granularity mining by adopting a shared cross-modal granularity decoder and hard negative match mechanism.With standard in-domain evaluation, CFAM establishes competitive performance across various datasets, especially on our ultra fine-grained UFine6926. Furthermore, by evaluating on UFine3C, we demonstrate that training on our UFine6926 significantly improves generalization to real scenarios compared with other coarse-grained datasets. The dataset and code will be made publicly available.

We conduct a comprehensive study on a new task named power battery detection (PBD), which aims to localize the dense cathode and anode plates endpoints from X-ray images to evaluate the quality of power batteries. Existing manufacturers usually rely on human eye observation to complete PBD, which makes it difficult to balance the accuracy and efficiency of detection. As the power source of new energy vehicles, the power battery is the most important system in the vehicle. It is crucial to strictly evaluate the quality of power batteries. However, there are no publicly available benchmark datasets and the AI-level baseline. Most factories rely on human eye observation and traditional image processing to complete PBD, which makes it difficult to ensure the accuracy and efficiency of detection at the same time.To address this issue and drive more attention into this meaningful task, we first elaborately collect a dataset, called X-ray PBD, which has $1,500$ diverse X-ray images selected from thousands of power batteries of $5$ manufacturers, with $7$ different visual interference. Then, we propose a novel segmentation-based solution for PBD, termed multi-dimensional collaborative network (MDCNet). With the help of line and counting predictors, the representation of the point segmentation branch can be improved at both semantic and detail aspects.Besides, we design an effective distance-adaptive mask generation strategy, which can alleviate the visual challenge caused by the inconsistent distribution density of plates to provide MDCNet with stable supervision. In addition, we design eight metrics based on the discriminant criteria in real-life factories to comprehensively evaluate the detection performance of the model. Without any bells and whistles, our segmentation-based MDCNet consistently outperforms various other corner detection, crowd counting and general/tiny object detection-based solutions, making it a strong baseline that can help facilitate future research in PBD. Finally, we share some potential difficulties and works for future researches.

We present MM-AU, a novel dataset for Multi-Modal Accident video Understanding. MM-AU contains 11,727 in-the-wild ego-view accident videos, each with tempo- rally aligned text descriptions. We annotate over 2.23 mil- lion object boxes and 58,650 pairs of video-based accident reasons, covering 58 accident categories. MM-AU sup- ports various accident understanding tasks, particularly multimodal video diffusion to understand accident cause- effect chains for safe driving. With MM-AU, we present an Abductive accident Video understanding framework for Safe Driving perception (AdVersa-SD). AdVersa-SD per- forms video diffusion via an Object-Centric Video Diffu- sion (OAVD) method which is driven by an abductive CLIP model. This model involves a contrastive interaction loss to learn the pair co-occurrence of normal, near-accident, accident frames with the corresponding text descriptions, such as accident reasons, prevention advice, and accident categories. OAVD enforces the object region learning while fixing the content of the original frame background in video generation, to find the dominant objects for certain acci- dents. Extensive experiments verify the abductive ability of AdVersa-SD and the superiority of OAVD against the state- of-the-art diffusion models. Additionally, we provide care- ful benchmark evaluations for object detection and accident reason answering since AdVersa-SD relies on precise object and accident reason information.

Large-scale datasets have fueled recent advancements in AI-based autonomous vehicle research. However, these datasets are usually collected from a single vehicle's one-time pass of a certain location, lacking multiagent interactions or repeated traversals of the same place. Such information could lead to transformative enhancements in autonomous vehicles' perception, prediction, and planning capabilities. To bridge this gap, in collaboration with the self-driving company May Mobility, we present the MARS dataset which unifies scenarios that enable MultiAgent, multitraveRSal, and multimodal autonomous vehicle research. More specifically, MARS is collected with a fleet of autonomous vehicles driving within a certain geographical area. Each vehicle has its own route and different vehicles may appear at nearby locations. Each vehicle is equipped with a LiDAR and surround-view RGB cameras. We curate two subsets in MARS: one facilitates collaborative driving with multiple vehicles simultaneously present at the same location, and the other enables memory retrospection through asynchronous traversals of the same location by multiple vehicles. We conduct experiments in place recognition and neural reconstruction. More importantly, MARS introduces new research opportunities and challenges such as multitraversal 3D reconstruction, multiagent perception, and unsupervised object discovery. Our data and codes can be found at https://ai4ce.github.io/MARS/.

Surveillance videos are important for public security. However, current surveillance video tasks mainly focus on classifying and localizing anomalous events. Existing methods are limited to detecting and classifying the predefined events with unsatisfactory semantic understanding, although they have obtained considerable performance. To address this issue, we propose a new research direction of surveillance video-and-language understanding (VALU), and construct the first multimodal surveillance video dataset. We manually annotate the real-world surveillance dataset UCFCrime with fine-grained event content and timing. Our newly annotated dataset, UCA (UCF-Crime Annotation), contains 23,542 sentences, with an average length of 20 words, and its annotated videos are as long as 110.7 hours. Furthermore, we benchmark SOTA models for four multimodal tasks on this newly created dataset, which serve as new baselines for surveillance VALU. Through experiments, we find that mainstream models used in previously public datasets perform poorly onsurveillance video, demonstrating new challenges in surveillance VALU. We also conducted experiments on multimodal anomaly detection. These results demonstrate that our multimodal surveillance learning can improve the performance of anomaly detection. All the experiments highlight the necessity of constructing this dataset to advance surveillance AI.

The use of models that have been pre-trained on natural image datasets like ImageNet may face some limitations. First, this use may be restricted due to copyright and license on the training images, and privacy laws. Second, these datasets and models may incorporate societal and ethical biases. Formula-driven supervised learning (FDSL) enables model pre-training to circumvent these issues. This consists of generating a synthetic image dataset based on mathematical formulae and pre-training the model on it.In this work, we propose novel FDSL datasets based on Mandelbulb Variations. These datasets contain RGB images that are projections of colored objects deriving from the 3D Mandelbulb fractal. Pre-training ResNet-50 on one of our proposed datasets MandelbulbVAR-1k enables an average top-1 accuracy over target classification datasets that is at least 1\% higher than pre-training on existing FDSL datasets. With regard to anomaly detection on MVTec AD, pre-training the WideResNet-50 backbone on MandelbulbVAR-1k enables PatchCore to achieve 97.2\% average image-level AUROC. This is only 1.9\% lower than pre-training on ImageNet-1k (99.1\%) and 4.5\% higher than pre-training on the second-best performing FDSL dataset i.e. VisualAtom-1k (92.7\%). Regarding Vision Transformer (ViT) pre-training, another dataset that we propose and coin MandelbulbVAR-Hybrid-21k enables ViT-Base to achieve 82.2\% top-1 accuracy on ImageNet-1k, which is 0.4\% higher than pre-training on ImageNet-21k (81.8\%) and only 0.1\% lower than pre-training on VisualAtom-1k (82.3\%).

Being able to map the activities of others into one's own point of view is one fundamental human skill even from a very early age. Taking a step toward understanding this human ability, we introduce EgoExoLearn, a large-scale dataset that emulates the human demonstration following process, in which individuals record egocentric videos as they execute tasks guided by demonstration videos.Focusing on the potential applications of daily assistance and professional support, EgoExoLearn contains egocentric and demonstration video data spanning 120 hours captured in daily life scenarios and specialized laboratories. Along with the videos we record high-quality gaze data and provide detailed multimodal annotations, formulating a playground for modeling the human ability to bridge asynchronous procedural actions from different viewpoints.To this end, we present benchmarks such as cross-view association, cross-view action planning, and cross-view referenced skill assessment, along with detailed analysis. We expect EgoExoLearn can serve as an important resource for bridging the actions across views, thus paving the way for creating AI agents capable of seamlessly learning by observing humans in the real world. The dataset and benchmark codes are available at https://github.com/OpenGVLab/EgoExoLearn.

Understanding human social behaviour is crucial in computer vision and robotics. Micro-level observations like individual actions fall short, necessitating a comprehensive approach that considers individual behaviour, intra-group dynamics, and social group levels for a thorough understanding. To address dataset limitations, this paper introduces JRDB-Social, an extension of JRDB. Designed to fill gaps in human understanding across diverse indoor and outdoor social contexts, JRDB-Social provides annotations at three levels: individual attributes, intra-group interactions, and social group context. This dataset aims to enhance our grasp of human social dynamics for robotic applications. Utilizing the recent cutting-edge multi-modal large language models, we evaluated our benchmark to explore their capacity to decipher social human behaviour.

Image datasets are essential not only in validating existing methods in computer vision but also in developing new methods. Most existing image datasets focus on trichromatic intensity images to mimic human vision.However, polarization and spectrum, the wave properties of light that animals in harsh environments and with limited brain capacity often rely on, remain underrepresented in existing datasets. Although spectro-polarimetric datasets exist, these datasets have insufficient object diversity, limited illumination conditions, linear-only polarization data, and inadequate image count. Here, we introduce two spectro-polarimetric datasets: trichromatic Stokes images and hyperspectral Stokes images. This novel dataset encompass both linear and circular polarization; they introduce multiple spectral channels; and they feature a broad selection of real-world scenes. With our dataset in hand, we analyze the spectro-polarimetric image statistics, develop efficient representations of such high-dimensional data, and evaluate spectral dependency of shape-from-polarization methods. As such, the proposed dataset promises a foundation for data-driven spectro-polarimetric imaging and vision research. Dataset and code will be publicly available.

We introduce MatSynth, a dataset of $4,000+$ CC0 ultra-high resolution PBR materials. Materials are crucial components of virtual relightable assets, defining the interaction of light at the surface of geometries. Given their importance, significant research effort was dedicated to their representation, creation and acquisition. However, in the past 6 years, most research in material acquisiton or generation relied either on the same unique dataset, or on company-owned huge library of procedural materials. With this dataset we propose a significantly larger, more diverse, and higher resolution set of materials than previously publicly available. We carefully discuss the data collection process and demonstrate the benefits of this dataset on material acquisition and generation applications. The complete data further contains metadata with each material's origin, license, category, tags, creation method and, when available, descriptions and physical size, as well as 3M+ renderings of the augmented materials, in 1K, under various environment lightings.

Visual grounding refers to the process of associating natural language expressions with corresponding regions within an image. Existing benchmarks for visual grounding primarily operate within small-scale scenes with a few objects. Nevertheless, recent advances in imaging technology have enabled the acquisition of gigapixel-level images, providing high-resolution details in large-scale scenes containing numerous objects. To bridge this gap between imaging and computer vision benchmarks and make grounding more practically valuable, we introduce a novel dataset, named GigaGrounding, designed to challenge visual grounding models in gigapixel-level large-scale scenes. We extensively analyze and compare the dataset with existing benchmarks, demonstrating that GigaGrounding presents unique challenges such as large-scale scene understanding, gigapixel-level resolution, significant variations in object scales, and the "multi-hop expressions". Furthermore, we introduced a simple yet effiective grounding approach, which employs a "glance-to-zoom-in" paradigm and exhibits enhanced capabilities for addressing the GigaGrounding task. The dataset and our code will be made publicly available upon paper acceptance.

Vehicle-to-everything (V2X) is a popular topic in the field of Autonomous Driving in recent years. Vehicle-infrastructure cooperation (VIC) becomes one of the important research area. Due to the complexity of traffic conditions such as blind spots and occlusion, it greatly limits the perception capabilities of single-view roadside sensing systems. To further enhance the accuracy of roadside perception and provide better information to the vehicle side, in this paper, we constructed holographic intersections with various layouts to build a large-scale multi-sensor holographic vehicle-infrastructure cooperation dataset, called HoloVIC. Our dataset includes 3 different types of sensors (Camera, Lidar, Fisheye) and employs 4 sensor-layouts based on the different intersections. Each intersection is equipped with 6-18 sensors to capture synchronous data. While autonomous vehicles pass through these intersections for collecting VIC data. HoloVIC contains in total on 100k+ synchronous frames from different sensors. Additionally, we annotated 3D bounding boxes based on Camera, Fisheye, and Lidar. We also associate the IDs of the same objects across different devices and consecutive frames in sequence. Based on HoloVIC, we formulated four tasks to facilitate the development of related research. We also provide benchmarks for these tasks.

The vision and language generative models have been overgrown in recent years. For video generation, various open-sourced models and public-available services have been developed to generate high-quality videos. However, these methods often use a few metrics, e.g., FVD or IS, to evaluate the performance. We argue that it is hard to judge the large conditional generative models from the simple metrics since these models are often trained on very large datasets with multi-aspect abilities. Thus, we propose a novel framework and pipeline for exhaustively evaluating the performance of the generated videos. Our approach involves generating a diverse and comprehensive list of 700 prompts for text-to-video generation, which is based on an analysis of real-world user data and generated with the assistance of a large language model. Then, we evaluate the state-of-the-art video generative models on our carefully designed benchmark, in terms of visual qualities, content qualities, motion qualities, and text-video alignment with 17 well-selected objective metrics. To obtain the final leaderboard of the models, we further fit a series of coefficients to align the objective metrics to the users' opinions. Based on the proposed human alignment method, our final score shows a higher correlation than simply averaging the metrics, showing the effectiveness of the proposed evaluation method.

The task of online mapping is to predict a local map using current sensor observations, e.g. from lidar and camera, without relying on a pre-built map. State-of-the-art methods are based on supervised learning and are trained predominantly using two datasets: nuScenes and Argoverse 2. However, these datasets revisit the same geographic locations across training, validation, and test sets. Specifically, over $80$\% of nuScenes and $40$\% of Argoverse 2 validation and test samples are less than $5$ m from a training sample. At test time, the methods are thus evaluated more on how well they localize within a memorized implicit map built from the training data than on extrapolating to unseen locations. Naturally, this data leakage causes inflated performance numbers and we propose geographically disjoint data splits to reveal the true performance in unseen environments. Experimental results show that methods perform considerably worse, some dropping more than $45$ mAP, when trained and evaluated on proper data splits. Additionally, a reassessment of prior design choices reveals diverging conclusions from those based on the original split. Notably, the impact of lifting methods and the support from auxiliary tasks (e.g., depth supervision) on performance appears less substantial or follows a different trajectory than previously perceived.

We have witnessed significant progress in deep learning-based 3D vision, ranging from neural radiance field (NeRF) based 3D representation learning to applications in novel view synthesis (NVS). However, existing scene-level datasets for deep learning-based 3D vision, limited to either synthetic environments or a narrow selection of real-world scenes, are quite insufficient. This insufficiency not only hinders a comprehensive benchmark of existing methods but also caps what could be explored in deep learning-based 3D analysis. To address this critical gap, we present DL3DV-10K, a large-scale scene dataset featuring 51.2 million frames from 10,510 videos captured from 65 types of point-of-interest (POI) locations, covering both bounded and unbounded scenes, with different levels of reflection, transparency, and lighting. We conducted a comprehensive benchmark of recent NVS methods on DL3DV-10K, which revealed valuable insights for future research in NVS. In addition, we have obtained encouraging results in a pilot study to learn generalizable NeRF from DL3DV-10K, which manifests the necessity of a large-scale scene-level dataset to forge a path toward a foundation model for learning 3D representation. Our DL3DV-10K dataset, benchmark results, and models will be publicly accessible.

Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in various multimodal tasks. However, their potential in the medical domain remains largely unexplored. A significant challenge arises from the scarcity of diverse medical images spanning various modalities and anatomical regions, which is essential in real-world medical applications. To solve this problem, in this paper, we introduce OmniMedVQA, a novel comprehensive medical Visual Question Answering (VQA) benchmark. This benchmark is collected from 73 different medical datasets, including 12 different modalities and covering more than 20 distinct anatomical regions. Importantly, all images in this benchmark are sourced from authentic medical scenarios, ensuring alignment with the requirements of the medical field and suitability for evaluating LVLMs. Through our extensive experiments, we have found that existing LVLMs struggle to address these medical VQA problems effectively. Moreover, what surprises us is that medical-specialized LVLMs even exhibit inferior performance to those general-domain models, calling for a more versatile and robust LVLM in the biomedical field. The evaluation results not only reveal the current limitations of LVLM in understanding real medical images but also highlight our dataset's significance. Our code with dataset are available at https://github.com/OpenGVLab/Multi-Modality-Arena.

The robust generalization of models to rare, in-distribution (ID) samples drawn from the long tail of the training distribution and to out-of-training-distribution (OOD) samples is one of the major challenges of current deep learning methods. For image classification, this manifests in the existence of adversarial attacks, the performance drops on distorted images, and a lack of generalization to concepts such as sketches. The current understanding of generalization in neural networks is very limited, but some biases that differentiate models from human vision have been identified and might be causing these limitations. Consequently, several attempts with varying success have been made to reduce these biases during training to improve generalization. We take a step back and sanity-check these attempts. Fixing the architecture to the well-established ResNet-50, we perform a large-scale study on 48 ImageNet models obtained via different training methods to understand how and if these biases - including shape bias, spectral biases, and critical bands - interact with generalization.Our extensive study results reveal that contrary to previous findings, these biases are insufficient to accurately predict the generalization of a model holistically.We provide access to all checkpoints and evaluation code at https://github.com/paulgavrikov/biasesvsgeneralization/

With the rapid development of Multi-modal Large Language Models (MLLMs), a number of diagnostic benchmarks have recently emerged to evaluate the comprehension capabilities of these models. However, most benchmarks predominantly assess spatial understanding in the static image tasks, while overlooking temporal understanding in the dynamic video tasks. To alleviate this issue, we introduce a comprehensive Multi-modal Video understanding Benchmark, namely MVBench, which covers 20 challenging video tasks that can not be effectively solved with a single frame. Specifically, we first introduce a novel static-to-dynamic method to define these temporal-related tasks. By transforming various static tasks into dynamic ones, we enable the systematic generation of video tasks that require a broad spectrum of temporal skills, ranging from perception to cognition. Then, guided by the task definition, we automatically convert public video annotations into multiple-choice QA to evaluate each task. On one hand, such a distinct paradigm allows us to build MVBench efficiently, without much manual intervention. On the other hand, it guarantees evaluation fairness with ground-truth video annotations, avoiding the biased scoring of LLMs. Moreover, we further develop a robust video MLLM baseline, i.e., MVChat, by progressive multi-modal training with diverse instruction-tuning data. The extensive results on our MVBench reveal that, the existing MLLMs are far from satisfactory in temporal understanding, while our MVChat largely surpasses these leading models by over 15% on MVBench. All models and data will be publicly available.

Recently, 3D anomaly detection, a crucial problem involving fine-grained geometry discrimination, is getting more attention. However, the lack of abundant real 3D anomaly data limits the scalability of current models. To enable scalable anomaly data collection, we propose a 3D anomaly synthesis pipeline to adapt existing large-scale 3D models for 3D anomaly detection. Specifically, we construct a synthetic dataset, i.e., Anomaly-ShapeNet, based on ShapeNet. Anomaly-ShapeNet consists of 1600 point cloud samples under 40 categories, which provides a rich and varied collection of data, enabling efficient training and enhancing adaptability to industrial scenarios. Meanwhile, to enable scalable representation learning for 3D anomaly localization, we propose a self-supervised method, i.e., Iterative Mask Reconstruction Network (IMRNet). During training, we propose a geometry-aware sample module to preserve potentially anomalous local regions during point cloud down-sampling. Then, we randomly mask out point patches and sent the visible patches to a transformer for reconstruction-based self-supervision. During testing, the point cloud repeatedly goes through the Mask Reconstruction Network, with each iteration’s output becoming the next input. By merging and contrasting the final reconstructed point cloud with the initial input, ourmethod successfully locates anomalies. Experiments show that IMRNet outperforms previous state-of-the-art methods, achieving 66.1% in I-AUC on our Anomaly-ShapeNetdataset and 72.5% in I-AUC on Real3D-AD dataset. Ourbenchmark will be released at https://github.com/Chopper233/Anomaly-ShapeNet.

Current state-of-the-art Video Object Segmentation (VOS) methods rely on dense per-object mask annotations both during training and testing. This requires time-consuming and costly video annotation mechanisms. We propose a novel Point-VOS task with a spatio-temporally sparse point-wise annotation scheme that substantially reduces the annotation effort. We apply our annotation scheme to two large-scale video datasets with text descriptions and annotate over 19M points across 133K objects in 32K videos. Based on our annotations, we propose a new Point-VOS benchmark, and a corresponding point-based training mechanism, which we use to establish strong baseline results. We show that existing VOS methods can easily be adapted to leverage our point annotations during training, and can achieve results close to the fully-supervised performance when trained on pseudo-masks generated from these points. In addition, we show that our data can be used to improve models that connect vision and language, by evaluating it on the Video Narrative Grounding (VNG) task. We will make our code and annotations available.

Despite recent advances in text-to-3D generative methods, there is a notable absence of reliable evaluation metrics. Existing metrics usually focus on a single criterion each, such as how well the asset aligned with the input text. These metrics lack the flexibility to generalize to different evaluation criteria and might not align well with human preferences. Conducting user preference studies is an alternative that offers both adaptability and human-aligned results. User studies, however, can be very expensive to scale. This paper presents an automatic, versatile, and human-aligned evaluation metric for text-to-3D generative models. To this end, we first develop a prompt generator using GPT-4V to generate evaluating prompts, which serve as input to compare text-to-3D models. We further design a method instructing GPT-4V to compare two 3D assets according to user-defined criteria. Finally, we use these pairwise comparison results to assign these models Elo ratings. Experimental results suggest our metric strongly align with human preference across different evaluation criteria. Our code is available at https://github.com/3DTopia/GPTEval3D .

While recent Vision-Language (VL) models excel at open-vocabulary tasks, it is unclear how to use them with specific or uncommon concepts. Personalized Text-to-Image Retrieval (TIR) or Generation (TIG) are recently introduced tasks that represent this challenge, where the VL model has to learn a concept from few images and respectively discriminate or generate images of the target concept in arbitrary contexts. We identify the ability to learn new meanings and their compositionality with known ones as two key properties of a personalized system. We show that the available benchmarks offer a limited validation of personalized textual concept learning from images with respect to the above properties and introduce ConCon-Chi as a benchmark for both personalized TIR and TIG, designed to fill this gap.We modelled the new-meaning concepts by crafting chimeric objects and formulating a large, varied set of contexts where we photographed each object. To promote the compositionality assessment of the learned concepts with known contexts, we combined different contexts with the same concept, and vice-versa. We carry out a thorough evaluation of state-of-the-art methods on the resulting dataset. Our study suggests that future work on personalized TIR and TIG methods should focus on the above key properties, and we propose principles and a dataset for their performance assessment. Dataset: https://doi.org/10.48557/QJ1166 and code: https://github.com/hsp-iit/concon-chi_benchmark.

Instance segmentation of neurons in volumetric light microscopy images of nervous systems enables groundbreaking research in neuroscience by facilitating joint functional and morphological analyses of neural circuits at cellular resolution. Yet said multi-neuron light microscopy data exhibits extremely challenging properties for the task of instance segmentation: Individual neurons have long-ranging, thin filamentous and widely branching morphologies, multiple neurons are tightly inter-weaved, and partial volume effects, uneven illumination and noise inherent to light microscopy severely impede local disentangling as well as long-range tracing of individual neurons. These properties reflect a current key challenge in machine learning research, namely to effectively capture long-range dependencies in the data. While respective methodological research is buzzing, to date methods are typically benchmarked on synthetic datasets. To address this gap, we release the FlyLight Instance Segmentation Benchmark (FISBe) dataset, the first publicly available multi-neuron light microscopy dataset with pixel-wise annotations. In addition, we define a set of instance segmentation metrics for benchmarking that we designed to be meaningful with regard to downstream analyses. Lastly, we provide three baselines to kick off a competition that we envision to both advance the field of machine learning regarding methodology for capturing long-range data dependencies, and facilitate scientific discovery in basic neuroscience. Project page: https://kainmueller-lab.github.io/fisbe

The analysis of the ubiquitous human-human interactions is pivotal for understanding humans as social beings. Existing human-human interaction datasets typically suffer from inaccurate body motions, lack of hand gestures and fine-grained textual descriptions. To better perceive and generate human-human interactions, we propose Inter-X, a currently largest human-human interaction dataset with accurate body movements and diverse interaction patterns, together with detailed hand gestures. The dataset includes 11K interaction sequences and more than 8.1M frames. We also equip Inter-X with versatile annotations of more than 34K fine-grained human part-level textual descriptions, semantic interaction categories, interaction order, and the relationship and personality of the subjects.Based on the elaborate annotations, we propose a unified benchmark composed of 4 categories of downstream tasks from both the perceptual and generative directions. Extensive experiments and comprehensive analysis show that Inter-X serves as a testbed for promoting the development of versatile human-human interaction analysis. Our dataset and benchmark will be publicly available for research purposes.

Acquiring large-scale, well-annotated datasets is essential for training robust scene text detectors, yet the process is often resource-intensive and time-consuming. While some efforts have been made to explore the synthesis of scene text images, a notable gap remains between synthetic and authentic data. In this paper, we introduce a novel method that utilizes Neural Radiance Fields (NeRF) to model real-world scenes and emulate the data collection process by rendering images from diverse camera perspectives, enriching the variability and realism of the synthesized data. A semi-supervised learning framework is proposed to categorize semantic regions within 3D scenes, ensuring consistent labeling of text regions across various viewpoints. Our method also models the pose, and view-dependent appearance of text regions, thereby offering precise control over camera poses and significantly improving the realism of text insertion and editing within scenes. Employing our technique on real-world scenes has led to the creation of a novel scene text image dataset. Compared to other existing benchmarks, the proposed dataset is distinctive in providing not only standard annotations such as bounding boxes and transcriptions but also the information of 3D pose attributes for text regions, enabling a more detailed evaluation of the robustness of text detection algorithms. Through extensive experiments, we demonstrate the effectiveness of our proposed method in enhancing the performance of scene text detectors.

In typical medical image classification problems, labeled data is scarce while unlabeled data is more available. Semi-supervised learning and self-supervised learning are two different research directions that can improve accuracy by learning from extra unlabeled data. Recent methods from both directions have reported significant gains on traditional benchmarks. Yet past benchmarks do not focus on medical tasks and rarely compare self- and semi- methods together on an equal footing. Furthermore, past benchmarks often handle hyperparameter tuning suboptimally. First, they may not tune hyperparameters at all, leading to underfitting. Second, when tuning does occur, it often unrealistically uses a labeled validation set that is much larger than the training set. Therefore currently published rankings might not always corroborate with their practical utility This study contributes a systematic evaluation of self- and semi- methods with a unified experimental protocol intended to guide a practitioner with scarce overall labeled data and a limited compute budget. We answer two key questions: Can hyperparameter tuning be effective with realistic-sized validation sets? If so, when all methods are tuned well, which self- or semi-supervised methods achieve the best accuracy? Our study compares 13 representative semi- and self-supervised methods to strong labeled-set-only baselines on 4 medical datasets. From 20000+ GPU hours of computation, we provide valuable best practices to resource-constrained practitioners: hyperparameter tuning is effective, and the semi-supervised method known as MixMatch delivers the most reliable gains across 4 datasets.

Computer vision applications predict on digital images acquired by a camera from physical scenes through light. However, conventional robustness benchmarks rely on perturbations in digitized images, diverging from distribution shifts occurring in the image acquisition process. To bridge this gap, we introduce a new distribution shift dataset, ImageNet-ES, comprising variations in environmental and camera sensor factors by directly capturing 202k images with a real camera in a controllable testbed. With the new dataset, we evaluate out-of-distribution (OOD) detection and model robustness. We find that existing OOD detection methods do not cope with the covariate shifts in ImageNet-ES, implying that the definition and detection of OOD should be revisited to embrace real-world distribution shifts. We also observe that the model becomes more robust in both ImageNet-C and -ES by learning environment and sensor variations in addition to existing digital augmentations. Lastly, our results suggest that effective shift mitigation via camera sensor control can significantly improve performance without increasing model size. With these findings, our benchmark may aid future research on robustness, OOD, and camera sensor control for computer vision. Our code and dataset are available at https://github.com/Edw2n/ImageNet-ES.

Perception plays a crucial role in various robot applications. However, existing well-annotated datasets are biased towards autonomous driving scenarios, while unlabelled SLAM datasets are quickly over-fitted, and often lack environment and domain variations. To expand the frontier of these fields, we introduce a comprehensive dataset named MCD (Multi-Campus Dataset), featuring a wide range of sensing modalities, high-accuracy ground truth, and diverse challenging environments across three Eurasian university campuses. MCD comprises both CCS (Classical Cylindrical Spinning) and NRE (Non-Repetitive Epicyclic) lidars, high-quality IMUs (Inertial Measurement Units), cameras, and UWB (Ultra-WideBand) sensors. Furthermore, in a pioneering effort, we introduce semantic annotations of 29 classes over 59k sparse NRE lidar scans across three domains, thus providing a novel challenge to existing semantic segmentation research upon this largely unexplored lidar modality. Finally, we propose, for the first time to the best of our knowledge, continuous-time ground truth based on optimization-based registration of lidar-inertial data on large survey-grade prior maps, which are also publicly released, each several times the size of existing ones. We conduct a rigorous evaluation of numerous state-of-the-art algorithms on MCD, report their performance, and highlight the challenges awaiting solutions from the research community.

Portable 360$^\circ$ cameras are becoming a cheap and efficient tool to establish large visual databases. By capturing omnidirectional views of a scene, these cameras could expedite building environment models that are essential for visual localization. However, such an advantage is often overlooked due to the lack of valuable datasets. This paper introduces a new benchmark dataset, 360Loc, composed of 360$^\circ$ images with ground truth poses for visual localization. We present a practical implementation of 360$^\circ$ mapping combining 360$^\circ$ images with lidar data to generate the ground truth 6DoF poses. 360Loc is the first dataset and benchmark that explores the challenge of cross-device visual positioning, involving 360$^\circ$ reference frames, and query frames from pinhole, ultra-wide FoV fisheye, and 360$^\circ$ cameras. We propose a virtual camera approach to generate lower-FoV query frames from 360$^\circ$ images, which ensures a fair comparison of performance among different query types in visual localization tasks. We also extend this virtual camera approach to feature matching-based and pose regression-based methods to alleviate the performance loss caused by the cross-device domain gap, and evaluate its effectiveness against state-of-the-art baselines. We demonstrate that omnidirectional visual localization is more robust in challenging large-scale scenes with symmetries and repetitive structures. These results provide new insights into 360-camera mapping and omnidirectional visual localization with cross-device queries. Project Page and dataset: https://huajianup.github.io/research/360Loc/.

In this paper, we propose Image Downscaling Assessment by Rate-Distortion (IDA-RD), a novel measure to quantitatively evaluate image downscaling algorithms. In contrast to image-based methods that measure the quality of downscaled images, ours is process-based that draws ideas from rate-distortion theory to measure the distortion incurred during downscaling. Our main idea is that downscaling and super-resolution (SR) can be viewed as the encoding and decoding processes in the rate-distortion model, respectively, and that a downscaling algorithm that preserves more details in the resulting low-resolution (LR) images should lead to less distorted high-resolution (HR) images in SR. In other words, the distortion should increase as the downscaling algorithm deteriorates. However, it is non-trivial to measure this distortion as it requires the SR algorithm to be blind and stochastic. Our key insight is that such requirements can be met by recent SR algorithms based on deep generative models that can find all matching HR images for a given LR image on their learned image manifolds.Extensive experimental results show the effectiveness of our IDA-RD measure.

Autonomous robot systems have attracted increasing research attention in recent years, where environment understanding is a crucial step for robot navigation, human-robot interaction, and decision-making. Real-world robot systems usually collect visual data from multiple sensors and are required to recognize numerous objects and their movements in complex human-crowded settings. Traditional benchmarks, with their reliance on single sensors and limited object classes and scenarios, fail to provide the comprehensive environmental understanding robots need for accurate navigation, interaction, and decision-making. As an extension of JRDB dataset, we unveil JRDB-PanoTrack, a novel open-world panoptic segmentation and tracking benchmark, towards more comprehensive environmental perception. JRDB-PanoTrack includes (1) various data involving indoor and outdoor crowded scenes, as well as comprehensive 2D and 3D synchronized data modalities; (2) high-quality 2D spatial panoptic segmentation and temporal tracking annotations, with additional 3D label projections for further spatial understanding; (3) diverse object classes for closed- and open-world recognition benchmarks, with OSPA-based metrics for evaluation. Extensive evaluation of leading methods shows significant challenges posed by our dataset. JRDB-PanoTrack is available at [hidden].

Multi-target multi-camera tracking is a crucial task that involves identifying and tracking individuals over time using video streams from multiple cameras. This task has practical applications in various fields, such as visual surveillance, crowd behavior analysis, and anomaly detection. However, due to the difficulty and cost of collecting and labeling data, existing datasets for this task are either synthetically generated or artificially constructed within a controlled camera network setting, which limits their ability to model real-world dynamics and generalize to diverse camera configurations. To address this issue, we present MTMMC, a real-world, large-scale dataset that includes long video sequences captured by 16 multi-modal cameras in two different environments - campus and factory - across various time, weather, and season conditions. This dataset provides a challenging test-bed for studying multi-camera tracking under diverse real-world complexities and includes an additional input modality of spatially aligned and temporally synchronized RGB and thermal cameras, which enhances the accuracy of multi-camera tracking. MTMMC is a super-set of existing datasets, benefiting independent fields such as person detection, re-identification, and multiple object tracking. We provide baselines and new learning setups on this dataset and set the reference scores for future studies. The datasets, models, and test server will be made publicly available.

The value of roadside perception, which could extend the boundaries of autonomous driving and traffic management, has gradually become more prominent and acknowledged in recent years. However, existing roadside perception approaches only focus on the single-infrastructure sensor system, which cannot realize a comprehensive understanding of a traffic area because of the limited sensing range and blind spots. Orienting high-quality roadside perception, we need Roadside Cooperative Perception (RCooper) to achieve practical area-coverage roadside perception for restricted traffic areas. Rcooper has its own domain-specific challenges, but further exploration is hindered due to the lack of datasets. We hence release the first real-world, large-scale RCooper dataset to bloom the research on practical roadside cooperative perception, including detection and tracking. The manually annotated dataset comprises 50k images and 30k point clouds, including two representative traffic scenes (i.e., intersection and corridor). The constructed benchmarks prove the effectiveness of roadside cooperation perception and demonstrate the direction of further research. Codes and dataset can be accessed at: https://github.com/AIR-THU/DAIR-RCooper.

Learning-based underwater image enhancement (UIE) methods have made great progress. However, the lack of large-scale and high-quality paired training samples has become the main bottleneck hindering the development of UIE. The inter-frame information in underwater videos can accelerate or optimize the UIE process. Thus, we constructed the first large-scale high-resolution underwater video enhancement benchmark (UVEB) to promote the development of underwater vision. UVEB is 500 times larger than the existing underwater image enhancement benchmark. It contains 1,308 pairs of video sequences and more than 453,000 high-resolution with 38\% Ultra-High-Definition (UHD) 4K frame pairs. UVEB comes from multiple countries, containing various scenes and video degradation types to adapt to diverse and complex underwater environments. We also propose the first supervised underwater video enhancement method, UVE-Net. UVE-Net converts the current frame information into convolutional kernels and passes them to adjacent frames for efficient inter-frame information exchange. By fully utilizing the redundant degraded information of underwater videos, UVE-Net completes video enhancement better. Experiments show the effective network design and good performance of UVE-Net.

The recently increased role of mobile photography has raised the standards of on-device photo processing tremendously. Despite the latest advancements in camera hardware, the mobile camera sensor area cannot be increased significantly due to physical constraints, leading to a pixel size of 0.6--2.0 $\mu$m, which results in strong image noise even in moderate lighting conditions. In the era of deep learning, one can train a CNN model to perform robust image denoising. However, there is still a lack of a substantially diverse dataset for this task. To address this problem, we introduce a novel Mobile Image Denoising Dataset (MIDD) comprising over 400,000 noisy / noise-free image pairs captured under various conditions by 20 different mobile camera sensors. Additionally, we propose a new DPreview test set consisting of data from 294 different cameras for precise model evaluation. Furthermore, we present the efficient baseline model SplitterNet for the considered mobile image denoising task that achieves high numerical and visual results, while being able to process 8MP photos directly on smartphone GPUs in under one second. Thereby outperforming models with similar runtimes. This model is also compatible with recent mobile NPUs, demonstrating an even higher speed when deployed on them. The conducted experiments demonstrate high robustness of the proposed solution when applied to images from previously unseen sensors, showing its high generalizability. The datasets, code and models can be found on the official project website

We introduce a new RGB-D object dataset captured in the wild called WildRGB-D. Unlike most existing real-world object-centric datasets which only come with RGB capturing, the direct capture of the depth channel allows better 3D annotations and broader downstream applications. WildRGB-D comprises large-scale category-level RGB-D object videos, which are taken using an iPhone to go around the objects in 360 degrees. It contains around 8500 recorded objects and nearly 20000 RGB-D videos across 46 common object categories. These videos are taken with diverse cluttered backgrounds with three setups to cover as many real-world scenarios as possible: (i) a single object in one video; (ii) multiple objects in one video; and (iii) an object with a static hand in one video. The dataset is annotated with object masks, real-world scale camera poses, and reconstructed aggregated point clouds from RGBD videos. We benchmark four tasks with WildRGB-D including novel view synthesis, camera pose estimation, object 6d pose estimation, and object surface reconstruction. Our experiments show that the large-scale capture of RGB-D objects provides a large potential to advance 3D object learning.

Retrieval tasks play central roles in real-world machine learning systems such as search engines, recommender systems, and retrieval-augmented generation (RAG). Achieving decent performance in these tasks often requires fine-tuning various pre-trained models on specific datasets and selecting the best candidate, a process that can be both time and resource-consuming. To tackle the problem, we introduce a novel and efficient method, called RetMMD, that leverages Maximum Mean Discrepancy (MMD) and kernel methods to assess the transferability of pre-trained models in retrieval tasks. RetMMD is calculated on pre-trained models and target datasets without any fine-tuning involved. Specifically, given some query, we quantify the distribution discrepancy between relevant and irrelevant document embeddings, by estimating the similarities within their mappings in the fine-tuned embedding space through kernel methods. This discrepancy is averaged over multiple queries, taking into account the distribution characteristics of the target dataset. Experiments suggest that the proposed metric calculated on pre-trained models closely aligns with retrieval performance post-fine-tuning. The observation holds across a variety of datasets, including image, text, and multi-modal domains, indicating the potential of using MMD and kernel methods for transfer learning evaluation in retrieval scenarios. In addition, we also design a way of evaluating dataset transferability for retrieval tasks, with experimental results demonstrating the effectiveness of the proposed approach.

The systematic evaluation and understanding of computer vision models under varying conditions require large amounts of data with comprehensive and customized labels, which real-world vision datasets rarely satisfy. While current synthetic data generators offer a promising alternative, particularly for embodied AI tasks, they often fall short for computer vision tasks due to low asset and rendering quality, limited diversity, and unrealistic physical properties. We introduce the BEHAVIOR Vision Suite (BVS), a set of tools and assets to generate fully customized synthetic data for systematic evaluation of computer vision models, based on the newly developed embodied AI benchmark, BEHAVIOR-1K. BVS supports a large number of adjustable parameters at the scene level (e.g., lighting, object placement), the object level (e.g., joint configuration, attributes such as "filled" and "folded"), and the camera level (e.g., field of view, focal length). Researchers can arbitrarily vary these parameters during data generation to perform controlled experiments. We showcase three example application scenarios: systematically evaluating the robustness of models across different continuous axes of domain shift, evaluating scene understanding models on the same set of images, and training and evaluating simulation-to-real transfer for a novel vision task: unary and binary state prediction. Project website: https://behavior-vision-suite.github.io/

Text-to-image generation has achieved astonishing results, yet precise spatial controllability and prompt fidelity remain highly challenging. This limitation is typically addressed through cumbersome prompt engineering, scene layout conditioning, or image editing techniques which often require hand drawn masks. Nonetheless, pre-existing works struggle to take advantage of the natural instance-level compositionality of scenes due to the typically flat nature of rasterized RGB output images. Towards adressing this challenge, we introduce MuLAn: a novel dataset comprising over 44K MUlti-Layer ANnotations of RGB images as multi-layer, instance-wise RGBA decompositions, and over 100K instance images. To build MuLAn, we developed a training free pipeline which decomposes a monocular RGB image into a stack of RGBA layers comprising of background and isolated instances. We achieve this through the use of pretrained general-purpose models, and by developing three modules: image decomposition for instance discovery and extraction, instance completion to reconstruct occluded areas, and image re-assembly. We use our pipeline to create MuLAn-COCO and MuLAn-LAION datasets, which contain a variety of image decompositions in terms of style, composition and complexity. With MuLAn, we provide the first photorealistic resource providing instance decomposition and occlusion information for high quality images, opening up new avenues for text-to-image generative AI research. With this, we aim to encourage the development of novel generation and editing technology, in particular layer-wise solutions. MuLAn data resources are available at https://MuLAn-dataset.github.io/.

Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning. We argue that this approach suffers from multiple limitations including: false positives and negatives due to CLIP's pretraining on noisy labels. We propose a pruning signal, Sieve, that employs synthetic captions generated by image-captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on unlabeled text corpus. Using DataComp, a multimodal dataset filtering benchmark, when evaluating on 38 downstream tasks, our pruning approach, surpasses CLIPScore by 2.6\% and 1.7\% on medium and large scale respectively. In addition, on retrieval tasks, Sieve leads to a significant improvement of 2.7\% and 4.5\% on medium and large scale respectively.

The increasing popularity of high dynamic range (HDR) imaging stems from its ability to faithfully capture luminance levels in natural scenes.However, HDR image quality assessment has been insufficiently addressed. Existing models are mostly designed for low dynamic range (LDR) images, which exhibit poorly correlated with human perception of HDR image quality. To fill this gap, we propose a family of HDR quality metrics by transferring the recent advancements in LDR domain. The key step in our approach is to employ a simple inverse display model to decompose an HDR image into a stack of LDR images with varying exposures. Subsequently, these LDR images are evaluated using state-of-the-art LDR quality metrics. Our family of HDR quality models offer three notable advantages. First, specific exposures (ie., luminance ranges) can be weighted to emphasize their assessment when calculating the overall quality score. Second, our HDR quality metrics directly inherit the capabilities of their base LDR quality models in assessing LDR images. Third, our metrics do not rely on human perceptual data of HDR image quality for re-calibration. Experiments conducted on four human-rated HDR image quality datasets indicate that our HDR quality metrics consistently outperform existing methods, including the HDR-VDP family. Furthermore, we demonstrate the promise of our models in the perceptual optimization of HDR novel view synthesis.

Large multimodal models (LMMs) have evolved from large language models (LLMs) to integrate multiple input modalities, such as visual inputs. This integration augments the capacity of LLMs for tasks requiring visual comprehension and reasoning. However, the extent and limitations of their enhanced abilities are not fully understood, especially when it comes to real-world tasks. To address this gap, we introduce GlitchBench, a novel benchmark derived from video game quality assurance tasks, to test and evaluate the reasoning capabilities of LMMs. Our benchmark is curated from a variety of unusual and glitched scenarios from video games and aims to challenge both the visual and linguistic reasoning powers of LMMs in detecting and interpreting out-of-the-ordinary events. We evaluate multiple state-of-the-art LMMs, and we show that GlitchBench presents a new challenge for these models. Code and data are available at: https://glitchbench.github.io/

We present WinSyn, a dataset of high resolution photographs and renderings of 3D models as a testbed for synthetic-to-real research. The dataset consists of 75,739 photographs of building windows, including traditional and modern designs, captured globally. Within these, we supply 89,318 crops containing windows, of which 9,002 are semantically labeled. Further, we present our domain-matched photorealistic procedural model which enables experimentation over a variety of parameter distributions and engineering approaches. Our procedural model provides a second corresponding dataset of 21,000 synthetic images. This jointly developed dataset is designed to facilitate research in the field of synthetic-to-real learning and synthetic data generation. WinSyn allows experimentation into the factors which make it challenging for synthetic data to complete with real world data. We perform ablations using our synthetic model to identify the salient rendering, materials, and geometric factors pertinent to accuracy within a labeling task. In addition, we leverage our dataset to explore the impact of semi-supervised approaches to synthetic modeling research.

Advances in neural fields are enabling high-fidelity capture of the shape and appearance of dynamic 3D scenes. However, their capabilities lag behind those offered by conventional representations such as 2D videos because of algorithmic challenges and the lack of large-scale multi-view real-world datasets. We address the dataset limitation with DiVa-360, a real-world 360$^\circ$ dynamic visual dataset that contains synchronized high-resolution and long-duration multi-view video sequences of table-scale scenes captured using a customized low-cost system with 53 cameras. It contains 21 object-centric sequences categorized by different motion types, 25 intricate hand-object interaction sequences, and 8 long-duration sequences for a total of 17.4 M image frames. In addition, we provide foreground-background segmentation masks, synchronized audio, and text descriptions. We benchmark the state-of-the-art dynamic neural field methods on DiVa-360 and provide insights about existing methods and future challenges on long-duration neural field capture.

Label noise, commonly found in real-world datasets, has a detrimental impact on a model's generalization. To effectively detect incorrectly labeled instances, previous works mostly rely on distinguishable training signals, e.g., training loss, as indicators to differentiate between clean and noisy labels. However, they have limitations in that the training signals incompletely reveal the model's behavior and are not effectively generalized to various noise types, resulting in limited detection accuracy. In this paper, we propose DynaCo framework that distinguishes incorrectly labeled instances from correctly labeled ones based on the dynamics of the training signals. To cope with the absence of supervision for clean and noisy labels, DynaCo first introduces a label corruption strategy that augments the original dataset with intentionally corrupted labels, enabling indirect simulation of the model's behavior on noisy labels. Then, DynaCo learns to identify clean and noisy instances by inducing two clearly distinguishable clusters from the latent representations of training dynamics. Our comprehensive experiments show that DynaCo outperforms the state-of-the-art competitors and shows strong robustness to various noise types and noise rates.

This paper presents DriveTrack, a new benchmark and data generation framework for long-range keypoint tracking in real-world videos. DriveTrack is motivated by the observation that the accuracy of state-of-the-art trackers depends strongly on visual attributes around the selected keypoints, such as texture and lighting. The problem is that these artifacts are especially pronounced in real-world videos, but these trackers are unable to train on such scenes due to a dearth of annotations. DriveTrack bridges this gap by building a framework to automatically annotate point tracks on autonomous driving datasets. We release a dataset consisting of 1 billion point tracks across 24 hours of video, which is seven orders of magnitude greater than prior real-world benchmarks and on par with the scale of synthetic benchmarks. DriveTrack unlocks new use cases for point tracking in real-world videos. First, we show that fine-tuning keypoint trackers on DriveTrack improves accuracy on real-world scenes by up to 7\%. Second, we analyze the sensitivity of trackers to visual artifacts in real scenes and motivate the idea of running assistive keypoint selectors alongside trackers.

Estimating 6D object poses is a major challenge in 3D computer vision. Building on successful instance-level approaches, research is shifting towards category-level pose estimation for practical applications. Current category-level datasets, however, fall short in annotation quality and pose variety. Addressing this, we introduce HouseCat6D, a new category-level 6D pose dataset. It features 1) multi-modality with Polarimetric RGB and Depth (RGBD+P), 2) encompasses 194 diverse objects across 10 household categories, including two photometrically challenging ones, and 3) provides high-quality pose annotations with an error range of only 1.35 mm to 1.74 mm. The dataset also includes 4) 41 large-scale scenes with comprehensive viewpoint and occlusion coverage, 5) a checkerboard-free environment, and 6) dense 6D parallel-jaw robotic grasp annotations. Additionally, we present benchmark results for leading category-level pose estimation networks.

When deploying segmentation models in practice, it is critical to evaluate their behaviors in varied and complex scenes.Different from the previous evaluation paradigms only in consideration of global attribute variations (e.g. adverse weather), we investigate both local and global attribute variations for robustness evaluation. To achieve this, we construct a mask-preserved attribute editing pipeline to edit visual attributes of real images with precise control of structural information. Therefore, the original segmentation labels can be reused for the edited images. Using our pipeline, we construct a benchmark covering both object and image attributes (e.g. color, material, pattern, style). We evaluate a broad variety of semantic segmentation models, spanning from conventional close-set models to recent open-vocabulary large models on their robustness to different types of variations. We find that both local and global attribute variations affect segmentation performances, and the sensitivity of models diverges across different variation types. We argue that local attributes have the same importance as global attributes, and should be considered in the robustness evaluation of segmentation models. Code: https://github.com/PRIS-CV/Pascal-EA

Recent advancements in large vision-language models enabled visual object detection in open-vocabulary scenarios, where object classes are defined in free-text formats during inference.In this paper, we aim to probe the state-of-the-art methods for open-vocabulary object detection to determine to what extent they understand fine-grained properties of objects and their parts.To this end, we introduce an evaluation protocol based on dynamic vocabulary generation to test whether models detect, discern, and assign the correct fine-grained description to objects in the presence of hard-negative classes.We contribute with a benchmark suite of increasing difficulty and probing different properties like color, pattern, and material.We further enhance our investigation by evaluating several state-of-the-art open-vocabulary object detectors using the proposed protocol and find that most existing solutions, which shine in standard open-vocabulary benchmarks, struggle to accurately capture and distinguish finer object details.We conclude the paper by highlighting the limitations of current methodologies and exploring promising research directions to overcome the discovered drawbacks. Data and code are available at https://lorebianchi98.github.io/FG-OVD .

High-quality human reconstruction and photo-realistic rendering of a dynamic scene is a long-standing problem in computer vision and graphics. Despite considerable efforts invested in developing various capture systems and reconstruction algorithms, recent advancements still struggle with loose or oversized clothing and overly complex poses. In part, this is due to the challenges of acquiring high-quality human datasets. To facilitate the development of these fields, in this paper, we present DyMVHumans, a versatile human-centric dataset for high-fidelity reconstruction and rendering of dynamic human scenarios from dense multi-view videos. It comprises 8.2 million frames captured by more than 56 synchronized cameras across diverse scenarios. These sequences comprise 32 human subjects across 45 different scenarios, each with a high-detailed appearance and realistic human motion. Inspired by recent advancements in neural radiance field (NeRF)-based scene representations, we carefully set up an off-the-shelf framework that is easy to provide those state-of-the-art NeRF-based implementations and benchmark on DyMVHumans dataset. It is paving the way for various applications like fine-grained foreground/background decomposition, high-quality human reconstruction and photo-realistic novel view synthesis of a dynamic scene. Extensive studies are performed on the benchmark, demonstrating new observations and challenges that emerge from using such high-fidelity dynamic data. DyMVHumans will be made publicly available to the community.

The boundless possibility of neural networks which can be used to solve a problem -- each with different performance -- leads to a situation where a Deep Learning (DL) expert is required to identify the best neural network. This goes against the hopes for DL to remove the need for experts. Neural Architecture Search (NAS) offers a solution for this by automatically identifying the best architecture. However, to date, NAS work has focused on a small set of datasets which we argue are not representative of real-world problems. We introduce eight new datasets that were created for a series of NAS Challenges (More details will be provided post review to maintain anonymity); AddNIST, Language, MultNIST, CIFARTile, Gutenberg, Isabella, GeoClassing, and Chesseract. The datasets and the challenges they were a part of were developed to direct attention to issues in NAS development and to encourage authors to consider how their models will perform on datasets unknown to them at development time. We present experimentation using standard Deep Learning methods as well as the best results from the participants of the challenge.

Parkinson's disease (PD) is a devastating movement disorder accelerating in prevalence globally, but the lack of precision symptom measurement has made it difficult to develop effective new therapies. The Unified Parkinson’s Disease Rating Scale (UPDRS) is the gold-standard label for assessing motor symptom severity, yet its scoring criteria are vague and subjective, resulting in coarse and noisy clinical assessments. While machine learning approaches could help to modernize PD symptom assessments by making them more quantitative, objective, and scalable, model development is hindered by the absence of publicly available video datasets for PD motor exams. Here, we introduce the TULIP dataset to bridge this gap. TULIP emphasizes precision and comprehensiveness, comprising multi-view video recordings (6 cameras) of all 25 UPDRS motor exam components, together with ratings by 3 clinical experts, in a cohort of Parkinson's patients and healthy controls. The multi-view recordings enable 3D reconstructions of body movement that better capture disease signatures than more conventional 2D methods. Using the dataset, we establish a baseline model for predicting UPDRS scores from 3D pose sequences, illustrating how existing diagnostics could be automated. Going forward, TULIP can be used to develop new precision diagnostics that transcend UPDRS scores to provide a deeper understanding of PD and its potential treatments.

Lithic Use-Wear Analysis (LUWA) using microscopic images is an underexplored vision-for-science research area. It seeks to distinguish the worked material, which is critical for understanding archaeological artifacts, material interactions, tool functionalities, and dental records. However, this challenging task goes beyond the well-studied image classification problem for common objects. It is affected by many confounders owing to the complex wear mechanism and microscopic imaging, which makes it difficult even for human experts to identify the worked material successfully. In this paper, we investigate the following three questions on this unique vision task for the first time:($\textbf{i}$) How well can state-of-the-art pre-trained models (like DINOv2) generalize to the rarely seen domain? ($\textbf{ii}$) How can few-shot learning be exploited for scarce microscopic images? ($\textbf{iii}$) How do the ambiguous magnification and sensing modality influence the classification accuracy? To study these, we collaborated with archaeologists and built the first open-source and the largest LUWA dataset containing 23,130 microscopic images with different magnifications and sensing modalities. Extensive experiments show that existing pre-trained models notably outperform human experts but still leave a large gap for improvements. Most importantly, the LUWA dataset provides an underexplored opportunity for vision and learning communities and complements existing image classification problems on common objects.

We introduce ShapeWalk, a carefully curated dataset designed to advance the field of language-guided compositional shape editing.The dataset consists of 158K unique shapes connected through 26K edit chains, with an average length of 14 chained shapes.Each consecutive pair of shapes is associated with precise language instructions describing the applied edits.We synthesize edit chains by reconstructing and interpolating shapes sampled from a realistic CAD-designed 3D dataset in the parameter space of a shape program.We leverage rule-based methods and language models to generate natural language prompts corresponding to each edit.To illustrate the practicality of our contribution, we train neural editor modules in the latent space of shape autoencoders, and demonstrate the ability of our dataset to enable a variety of language-guided shape edits.Finally, we introduce multi-step editing metrics to benchmark the capacity of our models to perform recursive shape edits.We hope that our work will enable further study of compositional language-guided shape editing, and finds application in 3D CAD design and interactive modeling.

Human perception of the world is shaped by a multitude of viewpoints and modalities. While many existing datasets focus on scene understanding from a certain perspective (e.g. egocentric or third-person views), our dataset offers a panoptic perspective (i.e. multiple viewpoints with multiple data modalities). Specifically, we encapsulate third-person panoramic and front views, as well as egocentric monocular/binocular views with rich modalities including video, multi-channel audio, directional audio, location data and textual scene descriptions within each scene captured, presenting comprehensive observation of the world. To the best of our knowledge, this is the first database that covers multiple viewpoints with multiple data modalities to mimic how daily information is accessed in the real world. Through our benchmark analysis, we presented 5 different scene understanding tasks on the proposed 360+x dataset to evaluate the impact and benefit of each data modality and perspective. Extensive experimental analysis reveals the effectiveness of each data modality and perspective in enhancing panoptic scene understanding. We hope the unique dataset could broaden the scope of comprehensive scene understanding and encourage the community to approach these problems from more diverse perspectives.

We present Ego-Exo4D, a diverse, large-scale multimodal multiview video dataset and benchmark challenge. Ego-Exo4D centers around simultaneously-captured egocentric and exocentric video of skilled human activities (e.g., sports, music, dance, bike repair). More than 800 participants from 13 cities worldwide performed these activities in 131 different natural scene contexts, yielding long-form captures from 1 to 42 minutes each and 1,422 hours of video combined. The multimodal nature of the dataset is unprecedented: the video is accompanied by multichannel audio, eye gaze, 3D point clouds, camera poses, IMU, and multiple paired language descriptions---including a novel ``expert commentary" done by coaches and teachers and tailored to the skilled-activity domain. To push the frontier of first-person video understanding of skilled human activity, we also present a suite of benchmark tasks and their annotations, including fine-grained activity understanding, proficiency estimation, cross-view translation, and 3D hand/body pose. All resources will be open sourced to fuel new research in the community.

Recent Text-to-Image (T2I) generation models such as Stable Diffusion and Imagen have made significant progress in generating high-resolution images based on text descriptions. However, many generated images still suffer from issues such as artifacts/implausibility, misalignment with text descriptions, and low aesthetic quality. Inspired by the success of Reinforcement Learning with Human Feedback (RLHF) for large language models, prior works collected human-provided scores as feedback on generated images and trained a reward model to improve the T2I generation. In this paper, we enrich the feedback signal by (i) marking image regions that are implausible or misaligned with the text, and (ii) annotating which words in the text prompt are misrepresented or missing on the image. We collect such rich human feedback on 18K generated images (RichHF-18K) and train a multimodal transformer to predict the rich feedback automatically. We show that the predicted rich human feedback can be leveraged to improve image generation, for example, by selecting high-quality training data to finetune and improve the generative models, or by creating masks with predicted heatmaps to inpaint the problematic regions. Notably, the improvements generalize to models (Muse) beyond those used to generate the images on which human feedback data were collected (Stable Diffusion variants). The RichHF-18K data set will be released in our GitHub repository: https://github.com/google-research/google-research/tree/master/richhf_18k.

Large multimodal language models have shown remarkable proficiency in understanding and editing images. However, a majority of these visually-tuned models struggle to comprehend the textual content embedded in images, primarily due to the limitation of training data. In this work, we introduce TRINS: a Text-Rich image\footnote{In this work, we use the phrase ``text-rich images'' to describe images with rich textual information, such as posters and book covers.} INStruction dataset, with the objective of enhancing the reading ability of the multimodal large language model. TRINS is built using hybrid data annotation strategies including machine-assisted and human-assisted annotation process. It contains 39,153 text-rich images, captions and 102,437 questions. Specifically, we show that the number of words per annotation in TRINS is significantly longer than that of related datasets, providing new challenges. Furthermore, we introduce a simple and effective architecture, called Language-vision Reading Assistant (LaRA), that is good at understanding textual contents within images. LaRA outperforms existing state-of-the-art multimodal large language models on the TRINS dataset as well as other classical benchmarks. Lastly, we conducted a comprehensive evaluation with TRINS on various text-rich image understanding and generation tasks, demonstrating its effectiveness.

We introduce MAGICK, a large-scale dataset of generated objects with high-quality alpha mattes. While image generation methods have produced segmentations, they cannot generate alpha mattes with accurate details in hair, fur, and transparencies. This is likely due to the small size of current alpha matting datasets and the difficulty in obtaining ground-truth alpha. We propose a scalable method for synthesizing images of objects with high-quality alpha that can be used as a ground-truth dataset. A key idea is to generate objects on a single-colored background so chroma keying approaches can be used to extract the alpha. However, this faces several challenges, including that current text-to-image generation methods cannot create images that can be easily chroma keyed and that chroma keying is an underconstrained problem that generally requires manual intervention for high-quality results. We address this using a combination of generation and alpha extraction methods. Using our method, we generate a dataset of 150,000 objects with alpha. We show the utility of our dataset by training an alpha-to-rgb generation method that outperforms baselines. Our dataset will be released to the public upon publication.

The existing facial datasets, while having plentiful images at near frontal views, lack images with extreme head poses, leading to the downgraded performance of deep learning models when dealing with profile or pitched faces. This work aims to address this gap by introducing a novel dataset named Extreme Pose Face High-Quality Dataset (EFHQ), which includes a maximum of 450k high-quality images of faces at extreme poses. To produce such a massive dataset, we utilize a novel and meticulous dataset processing pipeline to curate two publicly available datasets, VFHQ and CelebV-HQ, which contain many high-resolution face videos captured in various settings. Our dataset can complement existing datasets on various facial-related tasks, such as facial synthesis with 2D/3D-aware GAN, diffusion-based text-to-image face generation, and face reenactment. Specifically, training with EFHQ helps models generalize well across diverse poses, significantly improving performance in scenarios involving extreme views, confirmed by extensive experiments. Additionally, we utilize EFHQ to define a challenging cross-view face verification benchmark, in which the performance of SOTA face recognition models drops 5-37\% compared to frontal-to-frontal scenarios, aiming to stimulate studies on face recognition under severe pose conditions in the wild.

Neural fields (NeFs) have recently emerged as a versatile method for modeling signals of various modalities, including images, shapes, and scenes. Subsequently, a number of works have explored the use of NeFs as representations for downstream tasks, e.g. classifying an image based on the parameters of a NeF that has been fit to it. However, the impact of the NeF hyperparameters on their quality as downstream representation is scarcely understood and remains largely unexplored. This is in part caused by the large amount of time required to fit datasets of neural fields.In this work, we propose a JAX-based library that leverages parallelization to enable fast optimization of large-scale NeF datasets, resulting in a significant speed-up. With this library, we perform a comprehensive study that investigates the effects of different hyperparameters --including initialization, network architecture, and optimization strategies-- on fitting NeFs for downstream tasks.Our study provides valuable insights on how to train NeFs and offers guidance for optimizing their effectiveness in downstream applications.Finally, based on the proposed library and our analysis, we propose Neural Field Arena, a benchmark consisting of neural field variants of popular vision datasets, including MNIST, CIFAR, variants of ImageNet, and ShapeNetv2. Our library and the Neural Field Arena will be open-sourced to introduce standardized benchmarking and promote further research on neural fields.

Images of the natural world, collected by a variety of cameras, from drones to individual phones, are increasingly abundant sources of biological information. There is an explosion of computational methods and tools, particularly computer vision, for extracting biologically relevant information from images for science and conservation. Yet most of these are bespoke approaches designed for a specific task and are not easily adaptable or extendable to new questions, contexts, and datasets. A vision model for general organismal biology questions on images is of timely need. To approach this, we curate and release TreeOfLife-10M, the largest and most diverse ML-ready dataset of biology images. We then develop BioCLIP, a foundation model for the tree of life, leveraging the unique properties of biology captured by TreeOfLife-10M, namely the abundance and variety of images of plants, animals, and fungi, together with the availability of rich structured biological knowledge. We rigorously benchmark our approach on diverse fine-grained biology classification tasks, and find that BioCLIP consistently and substantially outperforms existing baselines (by 17% to 20% absolute). Intrinsic evaluation reveals that BioCLIP has learned a hierarchical representation conforming to the tree of life, shedding light on its strong generalizability. All data, code, and models will be publicly released upon acceptance.

The ability to detect unfamiliar or unexpected images is essential for safe deployment of computer vision systems. In the context of classification, the task of detecting images outside of a model's training domain is known as out-of-distribution (OOD) detection. While there has been a growing research interest in developing post-hoc OOD detection methods, there has been comparably little discussion around how these methods perform when the underlying classifier is not trained on a clean, carefully curated dataset. In this work, we take a closer look at 20 state-of-the-art OOD detection methods in the (more realistic) scenario where the labels used to train the underlying classifier are unreliable (e.g. crowd-sourced or web-scraped labels). Extensive experiments across different datasets, noise types \& levels, architectures and checkpointing strategies provide insights into the effect of class label noise on OOD detection, and show that poor separation between incorrectly classified ID samples vs. OOD samples is an overlooked yet important limitation of existing methods. Code: https://github.com/glhr/ood-labelnoise

Event cameras, with their high temporal and dynamic range and minimal memory usage, have found applications in various fields. However, their potential in static traffic monitoring remains largely unexplored. To facilitate this exploration, we present eTraM - a first-of-its-kind, fully event-based traffic monitoring dataset. eTraM offers 10 hr of data from different traffic scenarios in various lighting and weather conditions, providing a comprehensive overview of real-world situations. Providing 2M bounding box annotations, it covers eight distinct classes of traffic participants, ranging from vehicles to pedestrians and micro-mobility. eTraM's utility has been assessed using state-of-the-art methods for traffic participant detection, including RVT, RED, and YOLOv8. We quantitatively evaluate the ability of event-based models to generalize on nighttime and unseen scenes. Our findings substantiate the compelling potential of leveraging event cameras for traffic monitoring, opening new avenues for research and application. eTraM is available at https://eventbasedvision.github.io/eTraM.

Simultaneous localization and mapping (SLAM) is a fundamental task for numerous applications such as autonomous navigation and exploration. Despite many SLAM datasets have been released, current SLAM solutions still struggle to have sustained and resilient performance. One major issue is the absence of high-quality datasets including diverse all-weather conditions and a reliable metric for assessing robustness. This limitation significantly restricts the scalability and generalizability of SLAM technologies, impacting their development, validation, and deployment. To address this problem, we present SubT-MRS, an extremely challenging real-world dataset designed to push SLAM towards all-weather environments to pursue the most robust SLAM performance. It contains multi-degraded environments including over 30 diverse scenes such as structureless corridors, varying lighting conditions, and perceptual obscurants like smoke and dust; multimodal sensors such as LiDAR, fisheye camera, IMU, and thermal camera; and multiple locomotions like aerial, legged, and wheeled robots. We developed accuracy and robustness evaluation tracks for SLAM and introduced a novel robustness metric. Comprehensive studies are performed, revealing new observations, challenges, and opportunities for future research.

Public datasets, such as KITTI, nuScenes, and Waymo, have played a key role in the research and development of autonomous vehicles and advanced driver assistance systems. However, many of these datasets fail to incorporate a full range of driving conditions; some datasets only contain clear-weather conditions, underrepresenting or entirely missing colder weather conditions such as snow or autumn scenes with bright colorful foliage. In this paper, we present the Michigan State University Four Seasons (MSU-4S) Dataset, which contains real-world collections of autonomous vehicle data from varied types of driving scenarios. These scenarios were recorded throughout a full range of seasons, and capture clear, rainy, snowy, and fall weather conditions, at varying times of day. MSU-4S contains more than 100,000 two- and three-dimensional frames for camera, lidar, and radar data, as well as Global Navigation Satellite System (GNSS), wheel speed, and steering data, all annotated with weather, time-of-day, and time-of-year. Our data includes cluttered scenes that have large numbers of vehicles and pedestrians; and it also captures industrial scenes, busy traffic thoroughfare with traffic lights and numerous signs, and scenes with dense foliage. While providing a diverse set of scenes, our data incorporate an important feature: virtually every scene and its corresponding lidar, camera, and radar frames were captured in four different seasons, enabling unparalleled object detection analysis and testing of the domain shift problem across weather conditions. In that context, we present detailed analyses for 3D and 2D object detection showing a strong domain shift effect among MSU-4S data segments collected across different conditions. MSU-4S will also enable advanced multimodal fusion research including different combinations of camera-lidar-radar fusion, which continues to be of strong interest for the computer vision, autonomous driving and ADAS development communities. The MSU-4S dataset is available online at https://egr.msu.edu/waves/msu4s.

Cooperative perception offers several benefits for enhancing the capabilities of autonomous vehicles and improving road safety. Using roadside sensors in addition to onboard sensors increases reliability and extends the sensor range. External sensors offer higher situational awareness for automated vehicles and prevent occlusions. We propose CoopDet3D, a cooperative multi-modal fusion model, and TUMTraf-V2X, a perception dataset, for the cooperative 3D object detection and tracking task. Our dataset contains 2,000 labeled point clouds and 5,000 labeled images from five roadside and four onboard sensors. It includes 30k 3D boxes with track IDs and precise GPS and IMU data. We labeled nine categories and covered occlusion scenarios with challenging driving maneuvers, like traffic violations, near-miss events, overtaking, and U-turns. Through multiple experiments, we show that our CoopDet3D camera-LiDAR fusion model achieves an increase of +14.36 3D mAP compared to a vehicle camera-LiDAR fusion model. Finally, we make our dataset, model, labeling tool, and dev-kit publicly available on our website: https://tum-traffic-dataset.github.io/tumtraf-v2x.

Despite the commercial abundance of UAVs, aerial data acquisition remains challenging, and the existing Asia and North America-centric open-source UAV datasets are small-scale or low-resolution and lack diversity in scene contextuality. Additionally, the color content of the scenes, solar zenith angle, and population density of different geographies influence the data diversity. These factors conjointly render suboptimal aerial-visual perception of the deep neural network (DNN) models trained primarily on the ground view data, including the open-world foundational models. To pave the way for a transformative era of aerial detection, we present Multiview Aerial Visual RECognition (MAVREC), a video dataset where we record synchronized scenes from different perspectives --- ground camera and drone-mounted camera. MAVREC consists of around 2.5 hours of industry-standard 2.7K resolution video sequences, more than 0.5 million frames, and 1.1 million annotated bounding boxes. This makes MAVREC the largest ground and aerial-view dataset, and the fourth largest among all drone-based datasets across all modalities and tasks. Through our extensive benchmarking on MAVREC, we recognize that augmenting object detectors with ground-view images from the corresponding geographical location is a superior pre-training strategy for aerial detection. Building on this strategy, we benchmark MAVREC with a curriculum-based semi-supervised object detection approach that leverages labeled (ground and aerial) and unlabeled (only aerial) images to enhance aerial detection.

6DoF Pose estimation has been gaining increased importance in vision for over a decade, however it does not yet meet the reliability and accuracy standards for mass deployment in industrial robotics. To this effect, we present the first dataset and evaluation method for the co-evaluation of cameras, HDR, and algorithms targeted at reliable, high-accuracy industrial automation. Specifically, we capture 2,300 physical scenes of 22 industrial parts covering a $1m\times 1m\times 0.5m$ working volume, resulting in over 100,000 distinct object views. Each scene is captured with 13 well-calibrated multi-modal cameras including polarization and high-resolution structured light. In terms of lighting, we capture each scene at 4 exposures and in 3 challenging lighting conditions ranging from 100 lux to 100,000 lux. We also present, validate, and analyze robot consistency, an evaluation method targeted at scalable, high accuracy evaluation. We hope that vision systems that succeed on this dataset will have direct industry impact.

Vision-language models (VLMs) are trained for thousands of GPU hours on massive web scrapes, but they are not trained on all data indiscriminately. For instance, the LAION public dataset retained only about 10\% of the total crawled data. In recent times, data curation has gained prominence with several works developing strategies to retain "high-quality" subsets of "raw" scraped data. However, these strategies are typically developed agnostic to the available compute for training. In this paper, we demonstrate that making filtering decisions independent of training compute is often suboptimal---well-curated data rapidly loses its utilitywhen repeated, eventually decreasing below the utility of "unseen" but "lower-quality" data. In fact, we show that even a model trained on $\textit{unfiltered common crawl}$ obtains higher accuracy than that trained on the LAION dataset post 40 or more repetitions.While past research in neural scaling laws has considered web data to be homogenous, real data is not.Our work bridges this important gap in the literature by developing scaling trends that characterize the "utility" of various data subsets, accounting for the diminishing utility of a data point at its "nth" repetition.Our key message is that data curation $\textit{can not}$ be agnostic of the total compute a model will be trained for. Based on our analysis, we propose FADU (Filter by Assessing Diminishing Utility) that curates the best possible pool for achieving top performance on Datacomp at various compute budgets, carving out a pareto-frontier for data curation.

Existing audio-visual segmentation (AVS) datasets typically focus on short-trimmed videos with only one pixel-map annotation for a per-second video clip. In contrast, for untrimmed videos, the sound duration, start- and end-sounding time positions, and visual deformation of audible objects vary significantly. Therefore, we observed that current AVS models trained on trimmed videos might struggle to segment sounding objects in long videos. To investigate the feasibility of grounding audible objects in videos along both temporal and spatial dimensions, we introduce the Long-Untrimmed Audio-Visual Segmentation dataset (LU-AVS), which includes precise frame-level annotations of sounding emission times and provides exhaustive mask annotations for all frames. Considering that pixel-level annotations are difficult to achieve in some complex scenes, we also provide the bounding boxes to indicate the sounding regions. Specifically, LU-AVS contains 62K mask annotations across 6K videos, and 73K bounding box annotations across 7K videos. Compared with the existing datasets, LU-AVS videos are on average 4$\sim$8 times longer, with the silent duration being 2$\sim$10 times greater. Furthermore, we try our best to adapt some baseline models that were originally designed for audio-visual-relevant tasks to examine the challenges of our newly curated LU-AVS dataset. Through comprehensive evaluation, we demonstrate the challenges of the LU-AVS datasets compared to the ones containing trimmed videos. Therefore, LU-AVS provides an ideal yet challenging platform for evaluating audio-visual segmentation and localization on untrimmed long videos.

Self-attention mechanism is the key of the Transformer but often criticized for its computation demands. Previous token pruning works motivate their methods from the view of computation redundancy but still need to load the full network and require same memory costs. This paper introduces a novel strategy that simplifies vision transformers and reduces computational load through the selective removal of non-essential attention layers, guided by entropy considerations. We identify that regarding the attention layer in bottom blocks, their subsequent MLP layers, i.e. two feed-forward layers, can elicit the same entropy quantity. Meanwhile, the accompanied MLPs are under-exploited since they exhibit smaller feature entropy compared to those MLPs in the top blocks. Therefore, we propose to integrate the uninformative attention layers into their subsequent counterparts by degenerating them into identical mapping, yielding only MLP in certain transformer blocks. Experimental results on ImageNet-1k show that the proposed method can remove 40% attention layer of Deit-B, improving throughput and memory bound without performance compromise.

Weakly Supervised Semantic Segmentation (WSSS) aims to learn the concept of segmentation using image-level class labels. Recent WSSS works have shown promising results by using the Segment Anything Model (SAM), a foundation model for segmentation, during the inference phase. However, we observe that these methods can still be vulnerable to the noise of class activation maps (CAMs) serving as initial seeds. As a remedy, this paper introduces From-SAM-to-CAMs (S2C), a novel WSSS framework that directly transfers the knowledge of SAM to the classifier during the training process, enhancing the quality of CAMs itself. S2C comprises SAM-segment Contrasting (SSC) and a CAM-based prompting module (CPM), which exploit SAM at the feature and logit levels, respectively. SSC performs prototype-based contrasting using SAM's automatic segmentation results. It constrains each feature to be close to the prototype of its segment and distant from prototypes of the others. Meanwhile, CPM extracts prompts from the CAM of each class and uses them to generate class-specific segmentation masks through SAM. The masks are aggregated into unified self-supervision based on the confidence score, designed to consider the reliability of both SAM and CAMs. S2C achieves a new state-of-the-art performance across all benchmarks, outperforming existing studies by significant margins. The code is available at https://github.com/sangrockEG/S2C.

Test-time adaptation (TTA) aims to adapt a pre-trained model to a new test domain without access to source data after deployment. Existing approaches typically rely on self-training with pseudo-labels since ground-truth cannot be obtained from test data. Although the quality of pseudo labels is important for stable and accurate long-term adaptation, it has not been previously addressed. In this work, we propose DPLOT, a simple yet effective TTA framework that consists of two components: (1) domain-specific block selection and (2) pseudo-label generation using paired-view images. Specifically, we select blocks that involve domain-specific feature extraction and train these blocks by entropy minimization. After blocks are adjusted for current test domain, we generate pseudo-labels by averaging given test images and corresponding flipped counterparts. By simply using flip augmentation, we prevent a decrease in the quality of the pseudo-labels, which can be caused by the domain gap resulting from strong augmentation. Our experimental results demonstrate that DPLOT outperforms previous TTA methods in CIFAR10-C, CIFAR100-C, and ImageNet-C benchmarks, reducing error by up to 5.4\%, 9.1\%, and 2.9\%, respectively. Moreover, we provide an extensive analysis to demonstrate effectiveness of our framework. Our code will be available upon publication.

Recently, the advancement of self-supervised learning techniques, like masked autoencoders (MAE), has greatly influenced visual representation learning for images and videos.Nevertheless, it is worth noting that the predominant approaches in existing masked image / video modeling rely excessively on resource-intensive vision transformers (ViTs) as the feature encoder.In this paper, we propose a new approach termed as \textbf{VideoMAC}, which combines video masked autoencoders with resource-friendly ConvNets.Specifically, VideoMAC employs symmetric masking on randomly sampled pairs of video frames.To prevent the issue of mask pattern dissipation, we utilize ConvNets which are implemented with sparse convolutional operators as encoders. Simultaneously, we present a simple yet effective masked video modeling (MVM) approach, a dual encoder architecture comprising an online encoder and an exponential moving average target encoder, aimed to facilitate inter-frame reconstruction consistency in videos.Additionally, we demonstrate that VideoMAC, empowering classical (ResNet) / modern (ConvNeXt) convolutional encoders to harness the benefits of MVM, outperforms ViT-based approaches on downstream tasks, including video object segmentation (+\textbf{5.2\%} / \textbf{6.4\%} $\mathcal{J}$\&$\mathcal{F}$), body part propagation (+\textbf{6.3\%} / \textbf{3.1\%} mIoU), and human pose tracking (+\textbf{10.2\%} / \textbf{11.1\%} PCK@0.1).

Several unsupervised image segmentation approaches have been proposed which eliminate the need for dense manually-annotated segmentation masks; current models separately handle either semantic segmentation (e.g., STEGO) or class-agnostic instance segmentation (e.g., CutLER), but not both (i.e., panoptic segmentation). We propose an Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks---instance, semantic and panoptic---using a novel unified framework. U2Seg generates pseudo semantic labels for these segmentation tasks via leveraging self-supervised models followed by clustering; each cluster represents different semantic and/or instance membership of pixels. We then self-train the model on these pseudo semantic labels, yielding substantial performance gains over specialized methods tailored to each task: a +2.6 AP$^{\text{box}}$ boost (vs. CutLER) in unsupervised instance segmentation on COCO and a +7.0 PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOStuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation, which has not been previously explored. U2Seg is also a strong pretrained model for few-shot segmentation, surpassing CutLER by +5.0 AP$^{\text{mask}}$ when trained on a low-data regime, e.g., only 1\% COCO labels. We hope our simple yet effective method can inspire more research on unsupervised universal image segmentation.

Existing approaches to unsupervised video instance segmentation typically rely on motion estimates and experience difficulties tracking small or divergent motions. We present VideoCutLER, a simple method for unsupervised multi-instance video segmentation without using motion-based learning signals like optical flow or training on natural videos. Our key insight is that using high-quality pseudo masks and a simple video synthesis method for model training is surprisingly sufficient to enable the resulting video model to effectively segment and track multiple instances across video frames. We show the first competitive unsupervised learning results on the challenging YouTubeVIS-2019 benchmark, achieving 50.7\% AP$^{\text{video}}_{50}$, surpassing the previous state-of-the-art by a large margin. VideoCutLER can also serve as a strong pretrained model for supervised video instance segmentation tasks, exceeding DINO by 15.9\% on YouTubeVIS-2019 in terms of AP$^{\text{video}}$.

3D-aware Generative Adversarial Networks (GANs) have shown remarkable progress in learning to generate multi-view-consistent images and 3D geometries of scenes from collections of 2D images via neural volume rendering. Yet, the significant memory and computational costs of dense sampling in volume rendering have forced 3D GANs to adopt patch-based training or employ low-resolution rendering with post-processing 2D super resolution, which sacrifices multiview consistency and the quality of resolved geometry. Consequently, 3D GANs have not yet been able to fully resolve the rich 3D geometry present in 2D images. In this work, we propose techniques to scale neural volume rendering to the much higher resolution of native 2D images, thereby resolving fine-grained 3D geometry with unprecedented detail. Our approach employs learning-based samplers for accelerating neural rendering for 3D GAN training using up to 5 times fewer depth samples. This enables us to explicitly "render every pixel" of the full-resolution image during training and inference without post-processing superresolution in 2D. Together with our strategy to learn high-quality surface geometry, our method synthesizes high-resolution 3D geometry and strictly view-consistent images while maintaining image quality on par with baselines relying on post-processing super resolution. We demonstrate state-of-the-art 3D gemetric quality on FFHQ and AFHQ, setting a new standard for unsupervised learning of 3D shapes in 3D GANs.

Unsupervised object-centric learning aims to decompose scenes into interpretable object entities, termed slots. Slot-based auto-encoders stand out as a prominent method for this task. Within them, crucial aspects include guiding the encoder to generate object-specific slots and ensuring the decoder utilizes them during reconstruction. This work introduces two novel techniques, (i) an attention-based self-training approach, which distills superior slot-based attention masks from the decoder to the encoder, enhancing object segmentation, and (ii) an innovative patch-order permutation strategy for autoregressive transformers that strengthens the role of slot vectors in reconstruction. The effectiveness of these strategies is showcased experimentally. The combined approach significantly surpasses prior slot-based autoencoder methods in unsupervised object segmentation, especially with complex real-world images. We provide the implementation code at https://github.com/gkakogeorgiou/spot.

Category-level 3D pose estimation is a fundamentally important problem in computer vision and robotics, e.g. for embodied agents or to train 3D generative models. However, so far methods that estimate the category-level object pose require either large amounts of human annotations, CAD models or input from RGB-D sensors. In contrast, we tackle the problem of learning to estimate the category-level 3D pose only from casually taken object-centric videos without human supervision. We propose a two-step pipeline: First, we introduce a multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching using reconstructed coarse meshes and DINOv2 features. In a second step, the canonical poses and reconstructed meshes enable us to train a model for 3D pose estimation from a single image. In particular, our model learns to estimate dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. We demonstrate that our method outperforms all baselines at the unsupervised alignment of object-centric videos by a large margin and provides faithful and robust predictions in-the-wild on the Pascal3D+ and ObjectNet3D datasets.

Facial Attribute Classification (FAC) holds substantial promise in widespread applications. However, FAC models trained by traditional methodologies can be unfair by exhibiting accuracy inconsistencies across varied data subpopulations. This unfairness is largely attributed to bias in data, where some spurious attributes (e.g., Male) statistically correlate with the target attribute (e.g., Smiling). Most of existing fairness-aware methods rely on the labels of spurious attributes, which may be unavailable in practice. This work proposes a novel, generation-based two-stage framework to train a fair FAC model on biased data without additional annotation. Initially, we identify the potential spurious attributes based on generative models. Notably, it enhances interpretability by explicitly showing the spurious attributes in image space. Following this, for each image, we first edit the spurious attributes with a random degree sampled from a uniform distribution, while keeping target attribute unchanged. Then we train a fair FAC model by fostering model invariance to these augmentation. Extensive experiments on three common datasets demonstrate the effectiveness of our method in promoting fairness in FAC without compromising accuracy. Codes will be available.

In noisy label learning, estimating noisy class posteriors plays a fundamental role for developing consistent classifiers, as it forms the basis for estimating clean class posteriors and the transition matrix. Existing methods typically learn noisy class posteriors by training a classification model with noisy labels. However, when labels are incorrect, these models may be misled to overemphasize the feature parts that do not reflect the instance characteristics, resulting in significant errors in estimating noisy class posteriors. To address this issue, this paper proposes to augment the supervised information with part-level labels, encouraging the model to focus on and integrate richer information from various parts. Specifically, our method first partitions features into distinct parts by cropping instances, yielding part-level labels associated with these various parts. Subsequently, we introduce a novel single-to-multiple transition matrix to model the relationship between the noisy and part-level labels, which incorporates part-level labels into a classifier-consistent framework. Utilizing this framework with part-level labels, we can learn the noisy class posteriors more precisely by guiding the model to integrate information from various parts, ultimately improving the classification performance. Our method is theoretically sound, while experiments show that it is empirically effective in synthetic and real-world noisy benchmarks.

Unsupervised learning of keypoints and landmarks has seen significant progress with the help of modern neural network architectures, but performance is yet to match the supervised counterpart, making their practicability questionable. We leverage the emergent knowledge within text-to-image diffusion models, towards more robust unsupervised keypoints. Our core idea is to find text embeddings that would cause the generative model to consistently attend to compact regions in images (i.e. keypoints). To do so, we simply optimize the text embedding such that the cross-attention maps within the denoising network are localized as Gaussians with small standard deviations. We validate our performance on multiple dataset: the CelebA, CUB-200-2011, Tai-Chi-HD, DeepFashion, and Human3.6m datasets. We achieve significantly improved accuracy, sometimes even outperforming supervised ones, particularly for data that is non-aligned and less curated.

Images suffer from heavy spatial redundancy because pixels in neighboring regions are spatially correlated. Existing approaches strive to overcome this limitation by reducing less meaningful image regions. However, current leading methods rely on supervisory signals. They may compel models to preserve content that aligns with labeled categories and discard content belonging to unlabeled categories. This categorical inductive bias makes these methods less effective in real-world scenarios. To address this issue, we propose a self-supervised framework for image redundancy reduction called Learning to Rank Patches (LTRP). We observe that image reconstruction of masked image modeling models is sensitive to the removal of visible patches when the masking ratio is high (e.g., 90\%). Building upon it, we implement LTRP via two steps: inferring the semantic density score of each patch by quantifying variation between reconstructions with and without this patch, and learning to rank the patches with the pseudo score. The entire process is self-supervised, thus getting out of the dilemma of categorical inductive bias. We design extensive experiments on different datasets and tasks. The results demonstrate that LTRP outperforms both supervised and other self-supervised methods due to the fair assessment of image content. Code will be made available.

Federated learning achieves effective performance in modeling decentralized data. In practice, client data are not well-labeled, which makes it potential for federated unsupervised learning (FUSL) with non-IID data. However, the performance of existing FUSL methods suffers from insufficient representations, i.e., (1) representation collapse entanglement among local and global models, and (2) inconsistent representation spaces among local models. The former indicates that representation collapse in local model will subsequently impact the global model and other local models. The latter means that clients model data representation with inconsistent parameters due to the deficiency of supervision signals. In this work, we propose FedU2 which enhances generating uniform and unified representation in FUSL with non-IID data. Specifically, FedU2 consists of flexible uniform regularizer (FUR) and efficient unified aggregator (EUA). FUR in each client avoids representation collapse via dispersing samples uniformly, and EUA in server promotes unified representation by constraining consistent client model updating. To extensively validate the performance of FedU2, we conduct both cross-device and cross-silo evaluation experiments on two benchmark datasets, i.e., CIFAR10 and CIFAR100.

This paper proposes a GeneraLIst encoder-Decoder (GLID) pre-training method for better handling various downstream computer vision tasks. While self-supervised pre-training approaches, e.g., Masked Autoencoder, have shown success in transfer learning, task-specific sub-architectures are still required to be appended for different downstream tasks, which cannot enjoy the benefits of large-scale pre-training. GLID overcomes this challenge by allowing the pre-trained generalist encoder-decoder to be fine-tuned on various vision tasks with minimal task-specific architecture modifications. In the GLID training scheme, pre-training pretext task and other downstream tasks are modeled as "query-to-answer" problems, including the pre-training pretext task and other downstream tasks. We pre-train a task-agnostic encoder-decoder with query-mask pairs. During fine-tuning, GLID maintains the pre-trained encoder-decoder and queries, only replacing the topmost linear transformation layer with task-specific linear heads. This minimizes the pretrain-finetune architecture inconsistency and enables the pre-trained model to better adapt to downstream tasks. GLID achieves competitive performance on various vision tasks, including object detection, image segmentation, pose estimation, and depth estimation, outperforming or matching specialist models such as Mask2Former, DETR, ViTPose, and BinsFormer.

We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. Such pure vision models can possess capabilities for broad visual reasoning, analogous to those found in Large Language Models (LLMs.) To do this, we define a common format, "visual sentences", in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions without needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (420 billion tokens) is represented as sequences, the model can be trained to minimize cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable prompts at test time, showcasing remarkable generalization capabilities.

Self-Supervised Learning (SSL) has demonstrated promising results in 3D medical image analysis. However, the lack of high-level semantics in pre-training still heavily hinders the performance of downstream tasks. We observe that 3D medical images contain relatively consistent contextual position information, i.e., consistent geometric relations between different organs, which leads to a potential way for us to learn consistent semantic representations in pre-training. In this paper, we propose a simple-yet-effective Volume Contrast (VoCo) framework to leverage the contextual position priors for pre-training. Specifically, we first generate a group of base crops from different regions while enforcing feature discrepancy among them, where we employ them as class assignments of different regions. Then, we randomly crop sub-volumes and predict them belonging to which class (located at which region) by contrasting their similarity to different base crops, which can be seen as predicting contextual positions of different sub-volumes. Through this pretext task, VoCo implicitly encodes the contextual position priors into model representations without the guidance of annotations, enabling us to effectively improve the performance of downstream tasks that require high-level semantics. Extensive experimental results on six downstream tasks demonstrate the superior effectiveness of VoCo. Codes will be available.

Industrial anomaly detection (IAD) has garnered significant attention and experienced rapid development. However, the recent development of IAD approach has encountered certain difficulties due to dataset limitations. On the one hand, most of the state-of-the-art methods have achieved saturation (over 99\% in AUROC) on mainstream datasets such as MVTec, and the differences of methods cannot be well distinguished, leading to a significant gap between public datasets and actual application scenarios. On the other hand, the research on various new practical anomaly detection settings is limited by the scale of the dataset, posing a risk of overfitting in evaluation results. Therefore, we propose a large-scale, Real-world, and multi-view Industrial Anomaly Detection dataset, named Real-IAD, which contains 150K high-resolution images of 30 different objects, an order of magnitude larger than existing datasets. It has a larger range of defect area and ratio proportions, making it more challenging than previous datasets. To make the dataset closer to real application scenarios, we adopted a multi-view shooting method and proposed sample-level evaluation metrics. In addition, beyond the general unsupervised anomaly detection setting, we propose a new setting for Fully Unsupervised Industrial Anomaly Detection (FUIAD) based on the observation that the yield rate in industrial production is usually greater than 60\%, which has more practical application value. Finally, we report the results of popular IAD methods on the Real-IAD dataset, providing a highly challenging benchmark to promote the development of the IAD field.

Partial-label learning (PLL) is an important weakly supervised learning problem, which allows each training example to have a candidate label set instead of a single ground-truth label. Identification-based methods have been widely explored to tackle label ambiguity issues in PLL, which regard the true label as a latent variable to be identified. However, identifying the true labels accurately and completely remains challenging, causing noise in pseudo labels during model training. In this paper, we propose a new method called CroSel, which leverages historical prediction from models to identify true labels for most training examples. First, we introduce a cross selection strategy, which enables two deep models to select true labels of partially labeled data for each other. Besides, we propose a novel consistent regularization term called co-mix to avoid sample waste and tiny noise caused by false selection. In this way, CroSel can pick out the true labels of most examples with high precision. Extensive experiments demonstrate the superiority of CroSel, which consistently outperforms previous state-of-the-art methods on benchmark datasets. Additionally, our method achieves over 90\% accuracy and quantity for selecting true labels on CIFAR-type datasets under various settings.

Data mixing methods play a crucial role in semi-supervised learning (SSL), but their application is unexplored in long-tailed semi-supervised learning (LTSSL). The primary reason is that the in-batch mixing manner fails to address class imbalance. Furthermore, existing LTSSL methods mainly focus on re-balancing data quantity but ignore class-wise uncertainty, which is also vital for class balance. For instance, some classes with sufficient samples might still exhibit high uncertainty due to indistinguishable features. To this end, this paper introduces the Balanced and Entropy-based Mix (BEM), a pioneering mixing approach to re-balance the class distribution of both data quantity and uncertainty. Specifically, we first propose a class balanced mix bank to store data of each class for mixing. This bank samples data based on the estimated quantity distribution, thus re-balancing data quantity. Then, we present an entropy-based learning approach to re-balance class-wise uncertainty, including entropy-based sampling strategy, entropy-based selection module, and entropy-based class balanced loss. Our BEM first leverages data mixing for improving LTSSL, and it can also serve as a complement to the existing re-balancing methods. Experimental results show that BEM significantly enhances various LTSSL frameworks and achieves state-of-the-art performances across multiple benchmarks.

While recent model-free Reinforcement Learning (RL) methods have demonstrated human-level effectiveness in gaming environments, their success in everyday tasks like visual navigation has been limited, particularly under significant appearance variations. This limitation arises from (i) poor sample efficiency and (ii) over-fitting to training scenarios. To address these challenges, we present a world model that learns invariant features using (i) contrastive unsupervised learning and (ii) an intervention-invariant regularizer. Learning an explicit representation of the world dynamics i.e. a world model, improves sample efficiency while contrastive learning implicitly enforces learning of invariant features, which improves generalization. However, the na\"ive integration of contrastive loss to world models fails due to a lack of supervisory signals to the visual encoder, as world-model-based RL methods independently optimize representation learning and agent policy. To overcome this issue, we propose an intervention-invariant regularizer in the form of an auxiliary task such as depth prediction, image denoising, etc., that explicitly enforces invariance to style-interventions. Our method outperforms current state-of-the-art model-based and model-free RL methods and significantly on out-of-distribution point navigation task evaluated on the iGibson benchmark. We further demonstrate that our approach, with only visual observations, outperforms recent language-guided foundation models for point navigation, which is essential for deployment on robots with limited computation capabilities. Finally, we demonstrate that our proposed model excels at the sim-to-real transfer of its perception module on Gibson benchmark.

Novelty detection is a critical task for deploying machine learning models in the open world. A crucial property of novelty detection methods is universality, which can be interpreted as generalization across various distributions of training or test data. More precisely, for novelty detection, distribution shifts may occur in the training set or the test set. Shifts in the training set refer to cases where we train a novelty detector on a new dataset and expect strong transferability. Conversely, distribution shifts in the test set indicate the methods' performance when the trained model encounters a shifted test sample. We experimentally show that existing methods falter in maintaining universality, which stems from their rigid inductive biases. Motivated by this, we aim for more generalized techniques that have more adaptable inductive biases. In this context, we leverage the fact that contrastive learning provides an efficient framework to easily switch and adapt to new inductive biases through the proper choice of augmentations in forming the negative pairs. We propose a novel probabilistic auto-negative pair generation method AutoAugOOD, along with contrastive learning, to yield a universal novelty detector method. Our experiments demonstrate the superiority of our method under different distribution shifts in various image benchmark datasets. Notably, our method emerges universality in the lens of adaptability to different setups of novelty detection, including one-class, unlabeled multi-class, and labeled multi-class settings.

While recent supervised methods for reference-based object counting continue to improve the performance on benchmark datasets, they have to rely on small datasets due to the cost associated with manually annotating dozens of objects in images. We propose Unsupervised Counter (UnCo), a model that can learn this task without requiring any manual annotations. To this end, we construct "Self-Collages", images with various pasted objects as training samples, that provide a rich learning signal covering arbitrary object types and counts. Our method builds on existing unsupervised representations and segmentation techniques to successfully demonstrate for the first time the ability of reference-based counting without manual supervision. Our experiments show that our method not only outperforms simple baselines and generic models such as FasterRCNN and DETR, but also matches the performance of supervised counting models in some domains.

Pre-training a model and then fine-tuning it on downstream tasks has demonstrated significant success in the 2D image and NLP domains. However, due to the unordered and non-uniform density characteristics of point clouds, it is non-trivial to explore the prior knowledge of point clouds and pre-train a point cloud backbone. In this paper, we propose a novel pre-training method called Point cloud Diffusion pre-training (PointDif). We consider the point cloud pre-training task as a conditional point-to-point generation problem and introduce a conditional point generator. This generator aggregates the features extracted by the backbone and employs them as the condition to guide the point-to-point recovery from the noisy point cloud, thereby assisting the backbone in capturing both local and global geometric priors as well as the global point density distribution of the object. We also present a recurrent uniform sampling optimization strategy, which enables the model to uniformly recover from various noise levels and learn from balanced supervision. Our PointDif achieves substantial improvement across various real-world datasets for diverse downstream tasks such as classification, segmentation and detection. Specifically, PointDif attains 70.0% mIoU on S3DIS Area 5 for the segmentation task and achieves an average improvement of 2.4% on ScanObjectNN for the classification task compared to TAP. Furthermore, our pre-training framework can be flexibly applied to diverse point cloud backbones and bring considerable gains.

Learning representations to capture the very fundamental understanding of the world is a key challenge in machine learning.The hierarchical structure of explanatory factors hidden in data is such a general representation and could be potentially achieved witha hierarchical VAE through learning a hierarchy of increasingly abstract latent representations.However, training a hierarchical VAE always suffers from the "$\textit{posterior collapse}$'' issue, where the information of the input data is hard to propagate to the higher-level latent variables, hence resulting in a bad hierarchical representation.To address this issue, we first analyze the shortcomings of existing methods for mitigating the $\textit{posterior collapse}$ from an information theory perspective, then highlight the necessity of regularization for explicitly propagating data information to higher-level latent variables while maintaining the dependency between different levels.This naturally leads to formulating the inference of the hierarchical latent representation as a sequential decision process, which could benefit from applying reinforcement learning (RL) methodologies.To align RL's objective with the regularization, we first propose to employ a $\textit{skip-generative path}$ to acquire a reward for evaluating the information content of an inferred latent representation, and then the developed Q-value function based on it could have a consistent optimization direction of the regularization. Finally, policy gradient, one of the typical RL methods, is employed to train a hierarchical VAE without introducing a gradient estimator.Experimental results firmly support our analysis and demonstrate that our proposed method effectively mitigates the $\textit{posterior collapse}$ issue, learns an informative hierarchy, acquires explainable latent representations, and significantly outperforms other hierarchical VAE-based methods in downstream tasks.

Multi-view clustering (MVC) aims at exploring category structures among multi-view data in self-supervised manners. Multiple views provide more information than single views and thus existing MVC methods can achieve satisfactory performance. However, their performance might seriously degenerate when the views are noisy in practical multi-view scenarios. In this paper, we formally investigate the drawback of noisy views and then propose a theoretically grounded deep MVC method (namely MVCAN) to address this issue. Specifically, we propose a novel MVC objective that enables un-shared parameters and inconsistent clustering predictions across multiple views to reduce the side effects of noisy views. Furthermore, a two-level multi-view iterative optimization is designed to generate robust learning targets for refining individual views' representation learning. Theoretical analysis reveals that MVCAN works by achieving the multi-view consistency, complementarity, and noise robustness. Finally, experiments on extensive public datasets demonstrate that MVCAN outperforms state-of-the-art methods and is robust against the existence of noisy views.

Current self-supervised methods can primarily be categorized into contrastive learning and masked image modeling. Extensive studies have demonstrated that combining these two approaches can achieve state-of-the-art performance. However, these methods essentially reinforce the global consistency of contrastive learning without taking into account the conflicts between these two approaches, which hinders their generalizability to arbitrary scenarios. In this paper, we theoretically prove that MAE serves as a patch-level contrastive learning, where each patch within an image is considered as a distinct category. This presents a significant conflict with global-level contrastive learning, which treats all patches in an image as an identical category.To address this conflict, this work abandons the non-generalizable global-level constraints and proposes explicit patch-level contrastive learning as a solution. Specifically, this work employs the encoder of MAE to generate dual-branch features, which then perform patch-level learning through a decoder. In contrast to global-level data augmentation in contrastive learning, our approach leverages patch-level feature augmentation to mitigate interference from global-level learning. Consequently, our approach can learn heterogeneous representations from a single image while avoiding the conflicts encountered by previous methods. Massive experiments affirm the potential of our method for learning from arbitrary scenarios.

Establishing accurate 3D correspondences between shapes stands as a pivotal challenge with profound implications for computer vision and robotics. However, existing self-supervised methods for this problem assume perfect input shape alignment, restricting their real-world applicability. In this work, we introduce a novel self-supervised Rotation-Invariant 3D correspondence learner with Local Shape Transform, dubbed RIST, that learns to establish dense correspondences between shapes even under challenging intra-class variations and arbitrary orientations. Specifically, RIST learns to dynamically formulate an SO(3)-invariant local shape transform for each point, which maps the SO(3)-equivariant global shape descriptor of the input shape to a local shape descriptor. These local shape descriptors are provided as inputs to our decoder to facilitate point cloud self- and cross-reconstruction. Our proposed self-supervised training pipeline encourages semantically corresponding points from different shapes to be mapped to similar local shape descriptors, enabling RIST to establish dense point-wise correspondences. RIST demonstrates state-of-the-art performances on 3D part label transfer and semantic keypoint transfer given arbitrarily rotated point cloud pairs, outperforming existing methods by significant margins.

An important and unsolved problem in computer vision is to ensure that the algorithms are robust to changes in image domains. We address this problem in the scenario where we have access to images from the target domains but no annotations. Motivated by the challenges of the OOD-CV benchmark where we encounter real world Out-of-Domain (OOD) nuisances and occlusion, we introduce a novel Bayesian approach to OOD robustness for object classification. Our work extends Compositional Neural Networks (CompNets), which have been shown to be robust to occlusion but degrade badly when tested on OOD data. We exploit the fact that CompNets contain a generative head defined over feature vectors represented by von Mises-Fisher (vMF) kernels, which correspond roughly to object parts, and can be learned without supervision. We obverse that some vMF kernels are similar between different domains, while others are not. This enables us to learn a transitional dictionary of vMF kernels that are intermediate between the source and target domains and train the generative model on this dictionary using the annotations on the source domain, followed by iterative refinement. This approach, termed Unsupervised Generative Transition (UGT), performs very well in OOD scenarios even when occlusion is present. UGT is evaluated on different OOD benchmarks including the OOD-CV dataset, several popular datasets (e.g., ImageNet-C, artificial image corruptions (including adding occluders), and synthetic-to-real domain transfer, and does well in all scenarios outperforming SOTA alternatives (e.g. up to $10$% top-1 accuracy improvements on the Occluded OOD-CV benchmark).

Contrastive learning has emerged as a promising paradigm for 3D open-world understanding, i.e., aligning point cloud representation to image and text embedding space individually. In this paper, we introduce MixCon3D, a simple yet effective method aiming to sculpt holistic 3D representation in contrastive language-image-3D pre-training. In contrast to point cloud only, we develop the 3D object-level representation from complementary perspectives, e.g., multi-view rendered images with the point cloud. Then, MixCon3D performs language-3D contrastive learning, comprehensively depicting real-world 3D objects and bolstering text alignment. Additionally, we pioneer the first thorough investigation of various training recipes for the 3D contrastive learning paradigm, building a solid baseline with improved performance. Extensive experiments conducted on three representative benchmarks reveal that our method significantly improves over the baseline, surpassing the previous state-of-the-art performance on the challenging 1,156-category Objaverse-LVIS dataset by 5.7%. The versatility of MixCon3D is showcased in applications such as text-to-3D retrieval and point cloud captioning, further evidencing its efficacy in diverse scenarios. The code is available at https://github.com/UCSC-VLAA/MixCon3D.

Solving image and video jigsaw puzzles poses the challenging task of rearranging image fragments or video frames from unordered sequences to restore meaningful images and video sequences.Existing approaches often hinge on discriminative models tasked with predicting either the absolute positions of puzzle elements or the permutation actions applied to the original data. Unfortunately, these methods face limitations in effectively solving puzzles with a large number of elements.In this paper, we propose an innovative approach that harnesses diffusion transformers to address this challenge. Specifically, we generate positional information for image patches or video frames, conditioned on their underlying visual content. This information is then employed to accurately assemble the puzzle pieces in their correct positions, even in scenarios involving missing pieces.Our method achieves state-of-art performance in several datasets.

Implicit neural representations for video (NeRV) have recently become a novel way for high-quality video representation. However, existing works employ a single network to represent the entire video, which implicitly confuse static and dynamic information. This leads to an inability to effectively compress the redundant static information and lack the explicitly modeling of global temporal-coherent dynamic details. To solve above problems, we propose DS-NeRV, which decomposes videos into sparse learnable static codes and dynamic codes without the need for explicit optical flow or residual supervision. By setting different sampling rates for two codes and applying weighted sum and interpolation sampling methods, DS-NeRV efficiently utilizes redundant static information while maintaining high-frequency details. Additionally, we design a cross-channel attention-based (CCA) fusion module to efficiently fuse these two codes for frame decoding. Our approach achieves a high quality reconstruction of 31.2 PSNR with only 0.35M parameters thanks to separate static and dynamic codes representation and outperforms existing NeRV methods in many downstream tasks. Our project website is at https://haoyan14.github.io/DS-NeRV.

We developed a tool for visualizing and analyzing large pre-trained vision models by mapping them onto the brain, thus exposing their hidden inside. Our innovation arises from a surprising usage of brain encoding: predicting brain fMRI measurements in response to images. We report two findings. First, explicit mapping between the brain and deep-network features across dimensions of space, layers, scales, and channels is crucial. This mapping method, FactorTopy, is plug-and-play for any deep-network; with it, one can paint a picture of the network onto the brain (literally!). Second, our visualization shows how different training methods matter: they lead to remarkable differences in hierarchical organization and scaling behavior, growing with more data or network capacity. It also provides insight into finetuning: how pre-trained models change when adapting to small datasets. Our method is practical: only 3K images are enough to learn a network-to-brain mapping.

In pursuit of developing a robust ULD framework, we explore the potential of a recent paradigm of self-supervised learning algorithms, known as diffusion models. Some recent works have shown that these models implicitly contain important correspondence cues. Towards harnessing the potential of diffusion models for the ULD task, we make the following core contributions. First, we propose a ZeroShot ULD baseline based on simple clustering of random pixel locations with nearest neighbour matching. It delivers better results than existing ULD methods. Second, motivated by the ZeroShot performance, we develop a ULD algorithm based on diffusion features using self-training and clustering which also outperforms prior methods by notable margins. Third, we introduce a new proxy task based on generating latent pose codes and also propose a two-stage clustering to facilitate effective pseudo-labeling, resulting in a significant performance improvement. Overall, our approach consistently outperforms state-of-the-art methods on four challenging benchmarks AFLW, MAFL, CatHeads and LS3D by significant margins.

In the realm of point cloud scene understanding, particularly in indoor scenes, objects are arranged following human habits, resulting in objects of certain semantics being closely positioned and displaying notable inter-object correlations. This can create a tendency for neural networks to exploit these strong dependencies, bypassing the individual object patterns. To address this challenge, we introduce a novel self-supervised learning (SSL) strategy. Our approach leverages both object patterns and contextual cues to produce robust features. It begins with the formulation of an object-exchanging strategy, where pairs of objects with comparable sizes are exchanged across different scenes, effectively disentangling the strong contextual dependencies. Subsequently, we introduce a context-aware feature learning strategy, which encodes object patterns without relying on their specific context by aggregating object features across various scenes. Our extensive experiments demonstrate the superiority of our method over existing SSL techniques, further showing its better robustness to environmental changes. Moreover, we showcase the applicability of our approach by transferring pre-trained models to diverse point cloud datasets.

Object-centric learning (OCL) extracts the representation of objects with slots, offering an exceptional blend of flexibility and interpretability for abstracting low-level perceptual features. A widely adopted method within OCL is slot attention, which utilizes attention mechanisms to iteratively refine slot representations. However, a major drawback of most object-centric models, including slot attention, is their reliance on predefining the number of slots. This not only necessitates prior knowledge of the dataset but also overlooks the inherent variability in the number of objects present in each instance.To overcome this fundamental limitation, we present a novel complexity-aware object auto-encoder framework. Within this framework, we introduce an adaptive slot attention (AdaSlot) mechanism that dynamically determines the optimal number of slots based on the content of the data. This is achieved by proposing a discrete slot sampling module that is responsible for selecting an appropriate number of slots from a candidate list. Furthermore, we introduce a masked slot decoder that suppresses unselected slots during the decoding process.Our framework, tested extensively on object discovery tasks with various datasets, shows performance matching or exceeding top fixed-slot models. Moreover, our analysis substantiates that our method exhibits the capability to dynamically adapt the slot number according to each instance's complexity, offering the potential for further exploration in slot attention research. Project will be available at https://kfan21.github.io/AdaSlot/

Open-world Semi-Supervised Learning aims to classify unlabeled samples utilizing information from labeled data, while unlabeled samples are not only from the labeled known categories but also from novel categories previously unseen. Despite the promise, current approaches solely rely on hazardous similarity-based clustering algorithms and give unlabeled samples free rein to spontaneously group into distinct novel class clusters. Nevertheless, due to the absence of novel class supervision, these methods typically suffer from the representation collapse dilemma---features of different novel categories can get closely intertwined and indistinguishable, even collapsing into the same cluster and leading to degraded performance. To alleviate this, we propose a novel framework TRAILER which targets to attain an optimal feature arrangement revealed by the recently uncovered neural collapse phenomenon. To fulfill this, we adopt targeted prototypes that are pre-assigned uniformly with maximum separation and then progressively align the representations to them. To further tackle the potential downsides of such stringent alignment, we encapsulate a sample-target allocation mechanism with coarse-to-fine refinery that is able to infer label assignments with high quality. Extensive experiments demonstrate that TRAILER outperforms current state-of-the-art methods on generic and fine-grained benchmarks. The code is available at https://github.com/Justherozen/TRAILER.

We propose a hierarchical correlation clustering method that extends the well-known correlation clustering to produce hierarchical clusters applicable to both positive and negative pairwise dissimilarities. Then, in the following, we study unsupervised representation learning with such hierarchical correlation clustering. For this purpose, we first investigate embedding the respective hierarchy to be used for tree preserving embedding and feature extraction. Thereafter, we study the extension of minimax distance measures to correlation clustering, as another representation learning paradigm. Finally, we demonstrate the performance of our methods on several datasets.

We address the problem of generalized category discovery (GCD) that aims to partition a partially labeled collection of images; only a small part of the collection is labeled and the total number of target classes is unknown.To address this generalized image clustering problem, we revisit the mean-shift algorithm, i.e, a classic, powerful technique for mode seeking, and incorporate it into a contrastive representation learning framework. The proposed method, dubbed Contrastive Mean-Shift Learning, encourages the embedding network to learn image representations with better clustering properties by an iterative process of mean-shift and contrastive update.The proposed method introduces no extra learnable parts except the image feature extractor yet outperforms them on six public GCD benchmarks without bells and whistles.

In this paper, we introduce VoteCut, an innovative method for unsupervised object discovery that leverages feature representations from multiple self-supervised models. VoteCut employs normalized-cut based graph partitioning, clustering and a pixel voting approach. Additionally, We present CuVLER (Cut-Vote-and-LEaRn), a zero-shot model, trained using pseudo-labels, generated by VoteCut, and a novel soft target loss to refine segmentation accuracy. Through rigorous evaluations across multiple datasets and several unsupervised setups, our methods demonstrate significant improvements in comparison to previous state-of-the-art models. Our ablation studies further highlight the contributions of each component, revealing the robustness and efficacy of our approach. Collectively, VoteCut and CuVLER pave the way for future advancements in image segmentation.

We introduce SODA, a self-supervised diffusion model, designed for representation learning. The model incorporates an image encoder, which distills a source view into a compact representation, that, in turn, guides the generation of related novel views. We show that by imposing a tight bottleneck between the encoder and a denoising decoder, and leveraging novel view synthesis as a self-supervised objective, we can turn diffusion models into strong representation learners, capable of capturing visual semantics in an unsupervised manner. To the best of our knowledge, SODA is the first diffusion model to succeed at ImageNet linear-probe classification, and, at the same time, it accomplishes reconstruction, editing and synthesis tasks across a wide range of datasets. Further investigation reveals the disentangled nature of its emergent latent space, that serves as an effective interface to control and manipulate the produced images. All in all, we aim to shed light on the exciting and promising potential of diffusion models, not only for image generation, but also for learning rich and robust representations. See our website at soda-diffusion.github.io.

Event-based semantic segmentation has gained popularity due to its capability to deal with scenarios under high-speed motion and extreme lighting conditions, which cannot be addressed by conventional RGB cameras. Since it is hard to annotate event data, previous approaches rely on event-to-image reconstruction to obtain pseudo labels for training. However, this will inevitably introduce noise, and learning from noisy pseudo labels, especially when generated from a single source, may reinforce the errors. This drawback is also called confirmation bias in pseudo-labeling. In this paper, we propose a novel hybrid pseudo-labeling framework for unsupervised event-based semantic segmentation, HPL-ESS, to alleviate the influence of noisy pseudo labels. In particular, we first employ a plain unsupervised domain adaptation framework as our baseline, which can generate a set of pseudo labels through self-training. Then, we incorporate offline event-to-image reconstruction into the framework, and obtain another set of pseudo labels by predicting segmentation maps on the reconstructed images. A noisy label learning strategy is designed to mix the two sets of pseudo labels and enhance the quality. Moreover, we propose a soft prototypical alignment module to further improve the consistency of target domain features. Extensive experiments show that our proposed method outperforms existing state-of-the-art methods by a large margin on the DSEC-Semantic dataset (+5.88% accuracy, +10.32% mIoU), which even surpasses several supervised methods.

Positive-Unlabeled (PU) learning aims to train a binary classifier using minimal positive data supplemented by a substantially larger pool of unlabeled data, in the specific absence of explicitly annotated negatives. Despite its straightforward nature as a binary classification task, the currently best-performing PU algorithms still largely lag behind the supervised counterpart. In this work, we identify that the primary bottleneck lies in the difficulty of deriving discriminative representations under unreliable binary supervision with poor semantics, which subsequently hinders the common label disambiguation procedures. To cope with this problem, we propose a novel PU learning framework, namely Latent Group-Aware Meta Disambiguation (LaGAM), which incorporates a hierarchical contrastive learning module to extract the underlying grouping semantics within PU data and produce compact representations. As a result, LaGAM enables a more aggressive label disambiguation strategy, where we enhance the robustness of training by iteratively distilling the true labels of unlabeled data directly through meta-learning. Extensive experiments show that LaGAM significantly outperforms the current state-of-the-art methods by an average of 6.8\% accuracy on common benchmarks, approaching the supervised baseline. We also provide comprehensive ablations as well as visualized analysis to verify the effectiveness of our LaGAM.

Black-Box Knowledge Distillation (B2KD) is a formulated problem for cloud-to-edge model compression with invisible data and models hosted on the server. B2KD faces challenges such as limited Internet exchange and edge-cloud disparity of data distributions. In this paper, we formalize a two-step workflow consisting of deprivatization and distillation, and theoretically provide a new optimization direction from logits to cell boundary different from direct logits alignment. With its guidance, we propose a new method Mapping-Emulation KD (MEKD) that distills a black-box cumbersome model into a lightweight one. Our method does not differentiate between treating soft or hard responses, and consists of: 1) deprivatization: emulating the inverse mapping of the teacher function with a generator, and 2) distillation: aligning low-dimensional logits of the teacher and student models by reducing the distance of high-dimensional image points. For different teacher-student pairs, our method yields inspiring distillation performance on various benchmarks, and outperforms the previous state-of-the-art approaches.

Recent progress in self-supervised representation learning has resulted in models that are capable of extracting image features that are not only effective at encoding image-level, but also pixel-level, semantics. These features have been shown to be effective for dense visual semantic correspondence estimation, even outperforming fully-supervised methods. Nevertheless, current self-supervised approaches still fail in the presence of challenging image characteristics such as symmetries and repeated parts. To address these limitations, we propose a new approach for semantic correspondence estimation that supplements discriminative self-supervised features with 3D understanding via a weak geometric spherical prior. Compared to more involved 3D pipelines, our model only requires weak viewpoint information, and the simplicity of our spherical representation enables us to inject informative geometric priors into the model during training. We propose a new evaluation metric that better accounts for repeated part and symmetry-induced mistakes. We present results on the challenging SPair-71k dataset, where we show that our approach demonstrates is capable of distinguishing between symmetric views and repeated parts across many object categories, and also demonstrate that we can generalize to unseen classes on the AwA dataset.

We propose a self-supervised approach for learning physics-based subspaces for real-time simulation. Existing learning-based methods construct subspaces by approximating pre-defined simulation data in a purely geometric way. However, this approach tends to produce high-energy configurations, leads to entangled latent space dimensions, and generalizes poorly beyond the training set. To overcome these limitations, we propose a self-supervised approach that directly minimizes the system's mechanical energy during training. We show that our method leads to learned subspaces that reflect physical equilibrium constraints, resolve overfitting issues of previous methods, and offer interpretable latent space parameters.

Personalized Federated Learning (PFL) is proposed to find the greatest personalized models fitting to local data distribution for each client. To avoid the central failure and communication bottleneck in the server-based FL, we concentrate on the Decentralized Personalized Federated Learning (DPFL) that performs distributed model training in a Peer-to-Peer (P2P) manner. Most personalized works in DPFL are based on undirected topologies, in which clients communicate with each other symmetrically. However, the data, computation and communication resources heterogeneity result in large variances in the personalized models, which will lead the undirected and symmetric aggregation between such models to suboptimal personalized performance and unguaranteed convergence. To address these issues, we propose a directed collaboration DPFL framework by incorporating stochastic gradient push and partial model personalized, called $\textbf{D}ecentralized$ $\textbf{Fed}erated$ $\textbf{P}artial$ $\textbf{G}radient$ $\textbf{P}ush$ ($\textbf{DFedPGP}$). It personalizes the linear classifier in the modern deep model to customize the local solution for each client and learns a consensus representation in a fully decentralized manner. Clients only share gradients with a subset of neighbors based on the directed and asymmetric topologies, which guarantees flexible choices for resource efficiency and better convergence. Theoretically, we show that the proposed DFedPGP achieves a superior convergence rate of $\mathcal{O}(\frac{1}{\sqrt{T}})$ in the general non-convex setting, and tighter connectivity among clients will speed up the convergence. Extensive experiments demonstrate that the proposed method achieves stat-of-the-art (SOTA) accuracy in both data heterogeneity and computation resources heterogeneity scenarios, demonstrating the efficiency of the directed collaboration and partial gradient push.

Graph Contrastive Learning (GCL), a Self-Supervised Learning (SSL) architecture tailored for graphs, has shown notable potential for mitigating label scarcity. Its core idea is to amplify feature similarities between the positive sample pairs and reduce them between the negative sample pairs. Unfortunately, most existing GCLs consistently present suboptimal performances on both homophilic and heterophilic graphs. This is primarily attributed to two limitations of positive sampling, that is, incomplete local sampling and blind sampling. To address these limitations, this paper introduces a novel GCL framework with an adaptive positive sampling module, named grapH contrastivE Adaptive posiTive Samples (HEATS). Motivated by the observation that the affinity matrix corresponding to optimal positive sample sets has a block-diagonal structure with equal weights within each block, a self-expressive learning objective incorporating the block and idempotent constraint is presented. This learning objective and the contrastive learning objective are iteratively optimized to improve the adaptability and robustness of HEATS. Extensive experiments on graphs and images validate the effectiveness and generality of HEATS.

In the realm of computer vision and graphics, accurately establishing correspondences between geometric 3D shapes is pivotal for applications like object tracking, registration, texture transfer, and statistical shape analysis. Moving beyond traditional hand-crafted and data-driven feature learning methods, we incorporate spectral methods with deep learning, focusing on functional maps (FMs) and optimal transport (OT). Traditional OT-based approaches, often reliant on entropy regularization OT in learning-based framework, face computational challenges due to their quadratic cost. Our key contribution is to employ the sliced Wasserstein distance (SWD) for OT, which is a valid fast optimal transport metric in an unsupervised shape matching framework. This unsupervised framework integrates functional map regularizers with a novel OT-based loss derived from SWD, enhancing feature alignment between shapes treated as discrete probability measures. We also introduce an adaptive refinement process utilizing entropy regularized OT, further refining feature alignments for accurate point-to-point correspondences. Our method demonstrates superior performance in non-rigid shape matching, including near-isometric and non-isometric scenarios, and excels in downstream tasks like segmentation transfer. The empirical results on diverse datasets highlight our framework's effectiveness and generalization capabilities, setting new standards in non-rigid shape matching with efficient OT metrics and an adaptive refinement module.

Given a set of images, our goal is to map each image to a point in a feature space such that, not only point proximity indicates visual similarity, but where it is located directly encodes how prototypical the image is according to the dataset.Our key insight is to perform unsupervised feature learning in hyperbolic instead of Euclidean space, where the distance between points still reflects image similarity, yet we gain additional capacity for representing prototypicality with the location of the point: The closer it is to the origin, the more prototypical it is. The latter property is simply emergent from optimizing the metric learning objective: The image similar to many training instances is best placed at the center of corresponding points in Euclidean space, but closer to the origin in hyperbolic space.We propose an unsupervised feature learning algorithm in \underline{H}yperbolic space with sphere p\underline{ACK}ing. HACK first generates uniformly packed particles in the Poincar\'e ball of hyperbolic space and then assigns each image uniquely to a particle. With our feature mapper simply trained to spread out training instances in hyperbolic space, we observe that images move closer to the origin with congealing - a warping process that aligns all the images and makes them appear more common and similar to each other, validating our idea of unsupervised prototypicality discovery. We demonstrate that our data-driven prototypicality provides an easy and superior unsupervised instance selection to reduce sample complexity, increase model generalization with atypical instances and robustness with typical ones.

Vision-Language Models (VLMs) have demonstrated impressive performance on zero-shot classification, i.e. classification when provided merely with a list of class names. In this paper, we tackle the case of zero-shot classification in the presence of unlabeled data. We leverage the graph structure of the unlabeled data and introduce ZLaP, a method based on label propagation (LP) that utilizes geodesic distances for classification. We tailor LP to graphs containing both text and image features and further propose an efficient method for performing inductive inference based on a dual solution and a sparsification step. We perform extensive experiments to evaluate the effectiveness of our method on 14 common datasets and show that ZLaP outperforms the latest related works. Code: https://github.com/vladan-stojnic/ZLaP

Continual learning can empower vision-language models to continuously acquire new knowledge, without the need for access to the entire historical dataset. However, mitigating the performance degradation in large-scale models is non-trivial due to (i) parameter shifts throughout lifelong learning and (ii) significant computational burdens associated with full-model tuning. In this work, we present a parameter-efficient continual learning framework to alleviate long-term forgetting in incremental learning with vision-language models. Our approach involves the dynamic expansion of a pre-trained CLIP model, through the integration of Mixture-of-Experts (MoE) adapters in response to new tasks. To preserve the zero-shot recognition capability of vision-language models, we further introduce a Distribution Discriminative Auto-Selector (DDAS) that automatically routes in-distribution and out-of-distribution inputs to the MoE Adapter and the original CLIP, respectively. Through extensive experiments across various settings, our proposed method consistently outperforms previous state-of-the-art approaches while concurrently reducing parameter training burdens by 60%. Our code locates at https://github.com/JiazuoYu/MoE-Adapters4CL

Real-world systems often encounter new data over time, which leads to experiencing target domain shifts. Existing Test-Time Adaptation (TTA) methods tend to apply computationally heavy and memory-intensive backpropagation-based approaches to handle this. Here, we propose a novel method that uses a backpropagation-free approach for TTA for the specific case of 3D data. Our model uses a two-stream architecture to maintain knowledge about the source domain as well as complementary target-domain-specific information. The backpropagation-free property of our model helps address the well-known forgetting problem and mitigates the error accumulation issue. The proposed method also eliminates the need for the usually noisy process of pseudo-labeling and reliance on costly self-supervised training. Moreover, our method leverages subspace learning, effectively reducing the distribution variance between the two domains. Furthermore, the source-domain-specific and the target-domain-specific streams are aligned using a novel entropy-based adaptive fusion strategy. Extensive experiments on popular benchmarks demonstrate the effectiveness of our method. The code will be available at https://github.com/abie-e/BFTT3D.

Machine learning models struggle with generalization when encountering out-of-distribution (OOD) samples with unexpected distribution shifts. For vision tasks, recent studies have shown that test-time adaptation employing diffusion models can achieve state-of-the-art accuracy improvements on OOD samples by generating new samples that align with the model's domain without the need to modify the model's weights. Unfortunately, those studies have primarily focused on pixel-level corruptions, thereby lacking the generalization to adapt to a broader range of OOD types. We introduce Generalized Diffusion Adaptation (GDA), a novel diffusion-based test-time adaptation method robust against diverse OOD types. Specifically, GDA iteratively guides the diffusion by applying a marginal entropy loss derived from the model, in conjunction with style and content preservation losses during the reverse sampling process. In other words, GDA considers the model's output behavior with the semantic information of the samples as a whole, which can reduce ambiguity in downstream tasks during the generation process. Evaluation across various popular model architectures and OOD benchmarks shows that GDA consistently outperforms prior work on diffusion-driven adaptation. Notably, it achieves the highest classification accuracy improvements, ranging from 4.4\% to 5.02\% on ImageNet-C and 2.5\% to 7.4\% on Rendition, Sketch, and Stylized benchmarks. This performance highlights GDA's generalization to a broader range of OOD benchmarks.

Class-incremental learning (CIL) aims to enable models to continuously learn new classes while overcoming catastrophic forgetting. The introduction of pre-trained models has brought new tuning paradigms to CIL. In this paper, we revisit different parameter-efficient tuning (PET) methods within the context of continual learning. We observe that adapter tuning demonstrates superiority over prompt-based methods, even without parameter expansion in each learning session. Motivated by this, we propose incrementally tuning the shared adapter without imposing parameter update constraints, enhancing the learning capacity of the backbone. Additionally, we employ feature sampling from stored prototypes to retrain a unified classifier, further improving its performance. We estimate the semantic shift of old prototypes without access to past samples and update stored prototypes session by session. Our proposed method eliminates model expansion and avoids retaining any image samples. It surpasses previous pre-trained model-based CIL methods and demonstrates remarkable continual learning capabilities. Experimental results on five CIL benchmarks validate the effectiveness of our approach, achieving the state-of-the-art (SOTA) performance.

Even when using large multi-modal foundation models, few-shot learning is still challenging---if there is no proper inductive bias, it is nearly impossible to keep the nuanced class attributes while removing the visually prominent attributes that spuriously correlate with class labels. To this end, we find an inductive bias that the time-steps of a Diffusion Model (DM) can isolate the nuanced class attributes, i.e., as the forward diffusion adds noise to an image at each time-step, nuanced attributes are usually lost at an earlier time-step than the spurious attributes that are visually prominent. Building on this, we propose Time-step Few-shot (TiF) learner. We train class-specific low-rank adapters for a text-conditioned DM to make up for the lost attributes, such that images can be accurately reconstructed from their noisy ones given a prompt. Hence, at a small time-step, the adapter and prompt are essentially a parameterization of only the nuanced class attributes. For a test image, we can use the parameterization to only extract the nuanced class attributes for classification. TiF learner significantly outperforms OpenCLIP and its adapters on a variety of fine-grained and customized few-shot learning tasks. Codes are in Appendix.

Data-Free Meta-Learning (DFML) aims to extract knowledge from a collection of pre-trained models without requiring the original data, presenting practical benefits in contexts constrained by data privacy concerns. Current DFML methods primarily focus on the data recovery from these pre-trained models. However, they suffer from slow recovery speed and overlook gaps inherent in heterogeneous pre-trained models. In response to these challenges, we introduce the Faster and Better Data-Free Meta-Learning (FREE) framework, which contains: (i) a meta-generator for rapidly recovering training tasks from pre-trained models; and (ii) a meta-learner for generalizing to new unseen tasks. Specifically, within the module Faster Inversion via Meta-Generator, each pre-trained model is perceived as a distinct task. The meta-generator can rapidly adapt to a specific task in just five steps, significantly accelerating the data recovery. Furthermore, we propose Better Generalization via Meta-Learner and introduce an implicit gradient alignment algorithm to optimize the meta-learner. This is achieved as aligned gradient directions alleviate potential conflicts among tasks from heterogeneous pre-trained models. Empirical experiments on multiple benchmarks affirm the superiority of our approach, marking a notable speed-up (20$\times$) and performance enhancement (1.42\% $\sim$ 4.78\%) in comparison to the state-of-the-art.

In this paper, we present an empirical study on image recognition fairness, i.e., extreme class accuracy disparity on balanced data like ImageNet. We experimentally demonstrate that classes are not equal and the fairness issue is prevalent for image classification models across various datasets, network architectures, and model capacities. Moreover, several intriguing properties of fairness are identified. First, the unfairness lies in problematic representation rather than classifier bias. Second, with the proposed concept of \textit{Model Prediction Bias}, we investigate the origins of problematic representation during optimization. Our findings reveal that models tend to exhibit greater prediction biases for classes that are more challenging to recognize. It means that more other classes will be confused with harder classes. Then the False Positives (FPs) will dominate the learning in optimization, thus leading to their poor accuracy. Further, we conclude that data augmentation and representation learning algorithms improve overall performance by promoting fairness to some degree in image classification.

Low-shot counters estimate the number of objects corresponding to a selected category, based on only few or no exemplars annotated in the image. The current state-of-the-art estimates the total counts as the sum over the object location density map, but do not provide object locations and sizes, which are crucial for many applications. This is addressed by detection-based counters, which, howeverfall behind in the total count accuracy. Furthermore, both approaches tend to overestimate the counts in the presence of other object classes due to many false positives. We propose DAVE, a low-shot counter based on a detect-and-verify paradigm, that avoids the aforementioned issues by first generating a high-recall detection set and then verifying the detections to identify and remove the outliers.This jointly increases the recall and precision, leading to accurate counts. DAVE outperforms the top density-based counters by $\sim$20\% in the total count MAE, it outperforms the most recent detection-based counter by $\sim$20\% in detection quality, and sets a new state-of-the-art in zero-shot as well as text-prompt-based counting.

3D synthetic-to-real unsupervised domain adaptive segmentation is crucial to annotating new domains. Self-training is a competitive approach for this task, but its performance is limited by different sensor sampling patterns (i.e., variations in point density) and incomplete training strategies. In this work, we propose a density-guided translator (DGT), which translates point density between domains, and integrates it into a two-stage self-training pipeline named DGT-ST. First, in contrast to existing works that simultaneously conduct data generation and feature/output alignment within unstable adversarial training, we employ the non-learnable DGT to bridge the domain gap at the input level. Second, to provide a well-initialized model for self-training, we propose a category-level adversarial network in stage one that utilizes the prototype to prevent negative transfer. Finally, by leveraging the designs above, a domain-mixed self-training method with source-aware consistency loss is proposed in stage two to narrow the domain gap further. Experiments on two synthetic-to-real segmentation tasks (SynLiDAR $\rightarrow$ semanticKITTI and SynLiDAR $\rightarrow$ semanticPOSS) demonstrate that DGT-ST outperforms state-of-the-art methods, achieving 9.4$\%$ and 4.3$\%$ mIoU improvements, respectively. Our code will be released upon acceptance.

Domain adaptation for object detection typically entails transferring knowledge from one visible domain to another visible domain. However, there are limited studies on adapting from the visible to the thermal domain, because the domain gap between the visible and thermal domains is much larger than expected, and traditional domain adaptation can not successfully facilitate learning in this situation. To overcome this challenge, we propose a Distinctive Dual-Domain Teacher (D3T) framework that employs distinct training paradigms for each domain. Specifically, we segregate the source and target training sets for building dual-teachers and successively deploy exponential moving average to the student model to individual teachers of each domain. The framework further incorporates a zigzag learning method between dual teachers, facilitating a gradual transition from the visible to thermal domains during training. We validate the superiority of our method through newly designed experimental protocols with well-known thermal datasets, i.e., FLIR and KAIST. Code will be released publicly.

Recently, pre-trained vision-language models (e.g., CLIP) have shown great potential in few-shot learning and attracted a lot of research interest. Although many efforts have been made to improve few-shot generalization ability of CLIP, key factors on the effectiveness of existing methods have not been well studied, limiting further exploration of CLIP's potential in few-shot learning. In this paper, we first introduce a unified formulation to analyze CLIP-based few-shot learning methods from a perspective of logit bias, which encourages us to learn an effective logit bias for further improving performance of CLIP-based few-shot learning methods. To this end, we disassemble three key components involved in computation of logit bias (i.e., logit features, logit predictor, and logit fusion) and empirically analyze the effect on performance of few-shot classification. According to the analysis on the key components, this paper proposes a novel AMU-Tuning method to learn effective logit bias for CLIP-based few-shot classification. Specifically, our AMU-Tuning predicts logit bias by exploiting the appropriate $\underline{\textbf{A}}$uxiliary features, which are fed into an efficient feature-initialized linear classifier with $\underline{\textbf{M}}$ulti-branch training. Finally, an $\underline{\textbf{U}}$ncertainty-based fusion is developed to incorporate logit bias into CLIP for few-shot classification. The experiments are conducted on several widely used benchmarks, and the results show our proposed AMU-Tuning clearly outperforms its counterparts while achieving state-of-the-art performance without bells and whistles.

Universal Domain Adaptation (UniDA) targets knowledge transfer in the presence of both covariate and label shifts. Recently, Source-free Universal Domain Adaptation (SF-UniDA) has emerged to achieve UniDA without access to source data, which tends to be more practical due to data protection policies. The main challenge lies in determining whether covariate-shifted samples belong to target-private unknown categories. Existing methods tackle this either through hand-crafted thresholding or by developing time-consuming iterative clustering strategies. In this paper, we propose a new idea of LEArning Decomposition (LEAD), which decouples features into source-known and -unknown components to identify target-private data. Technically, LEAD initially leverages the Independent Component Analysis (ICA) for feature decomposition. Then, LEAD builds instance-level decision boundaries to adaptively identify target-private data. Extensive experiments across various UniDA scenarios have demonstrated the effectiveness and superiority of LEAD. Notably, in the OPDA scenario on VisDA dataset, LEAD outperforms GLC by 3.5% overall H-score and reduces 75% time to derive pseudo-labeling decision boundaries. Besides, LEAD is also appealing in that it is complementary to most methods. When integrated into UMAD, LEAD improves the baseline by 7.9% overall H-score in the OPDA scenario on Office-Home dataset.

Generalized Zero-Shot Learning (GZSL) methods often assume that the unseen classes are similar to seen classes, and thus perform poor when unseen classes are dissimilar to seen classes. Although some existing GZSL approaches can alleviate this issue by leveraging additional semantic information from test unseen classes, their generalization ability to dissimilar unseen classes is still unsatisfactory. This motivates us to study GZSL in the more practical setting, where unseen classes can be either similar or dissimilar to seen classes. In this paper, we propose a simple yet effective GZSL framework by exploring diverse semantics from external class names (DSECN), which is simultaneously robust on the similar and dissimilar unseen classes. This is achieved by introducing diverse semantics from external class names and aligning the introduced semantics to visual space using the classification head of pre-trained network. Furthermore, we show that the design idea of DSECN can easily be integrate into other advanced GZSL approaches, such as the generative-based ones, and enhance their robustness for dissimilar unseen classes. Extensive experiments in the practical setting including both similar and dissimilar unseen classes show that our method significantly outperforms the state-of-the-art approaches on all datasets and can be trained very efficiently.

It is a well-known fact that the performance of deep learning models deteriorates when they encounter a distribution shift at test-time. Test-Time Adaptation (TTA) algorithms have been proposed to adapt the model online while inferring test data. However, existing research predominantly focuses on classification tasks through the optimization of batch normalization layers or classification heads, but this approach limits its applicability to various model architectures like Transformers and makes it challenging to be applied to other tasks, such as object detection. In this paper, we propose a novel online adaption approach for object detection in continually changing test domains, considering which part of the model to update, how to update it, and when to perform the update. By introducing architecture-agnostic and light-weight adaptor modules and only updating these while leaving the pre-trained backbone unchanged, we can rapidly adapt to new test domains in an efficient way and prevent catastrophic forgetting. Furthermore, we present a practical and straightforward class-wise feature aligning method for object detection to resolve domain shifts. Additionally, we enhace efficiency by determining when the model is sufficiently adapted or when additional adaptation is needed due to changes in the test distribution. Our approach surpasses baselines on widely used benchmarks, achieving improvements of up to 4.9 and 7.9 in mAP for COCO $\rightarrow$ COCO-corrupted and SHIFT, respectively, while maintaining about 20 FPS or higher.

Large vision-language models (VLMs) like CLIP have demonstrated good zero-shot learning performance in the unsupervised domain adaptation task. Yet, most transfer approaches for VLMs focus on either the language or visual branches, overlooking the nuanced interplay between both modalities. In this work, we introduce a Unified Modality Separation (UniMoS) framework for unsupervised domain adaptation. Leveraging insights from modality gap studies, we craft a nimble modality separation network that distinctly disentangles CLIP's features into language-associated and vision-associated components. Our proposed Modality-Ensemble Training (MET) method fosters the exchange of modality-agnostic information while maintaining modality-specific nuances. We align features across domains using a modality discriminator. Comprehensive evaluations on three benchmarks reveal our approach sets a new state-of-the-art with minimal computational costs.

Conventional Unsupervised Domain Adaptation (UDA) strives to minimize distribution discrepancy between domains, which neglects to harness rich semantics from data and struggles to handle complex domain shifts. A promising technique is to leverage the knowledge of large-scale pre-trained vision-language models for more guided adaptation. Despite some endeavors, current methods often learn textual prompts to embed domain semantics for source and target domains separately and perform classification within each domain, limiting cross-domain knowledge transfer. Moreover, prompting only the language branch lacks flexibility to adapt both modalities dynamically. To bridge this gap, we propose Domain-Agnostic Mutual Prompting (DAMP) to exploit domain-invariant semantics by mutually aligning visual and textual embeddings. Specifically, the image contextual information is utilized to prompt the language model in a domain-agnostic and instance-conditioned way. Meanwhile, visual prompts are imposed based on the domain-agnostic textual prompt to elicit domain-invariant visual embeddings. These two branches of prompts are learned mutually with a cross-attention module and regularized with a semantic-consistency loss and an instance-discrimination contrastive loss. Experiments on three UDA benchmarks demonstrate the superiority of DAMP over state-of-the-art approaches.

The success of large language models has inspired the computer vision community to explore image segmentation foundation model that is able to zero/few-shot generalize through prompt engineering. Segment-Anything (SAM), among others, is the state-of-the-art image segmentation foundation model demonstrating strong zero/few-shot generalization. Despite the success, recent studies reveal the weakness of SAM under strong distribution shift. In particular, SAM performs awkwardly on corrupted natural images, camouflaged images, medical images, etc. Motivated by the observations, we aim to develop a self-training based strategy to adapt SAM to target distribution. Given the unique challenges of large source dataset, high computation cost and incorrect pseudo label, we propose a weakly supervised self-training architecture with anchor regularization and low-rank finetuning to improve the robustness and computation efficiency of adaptation. We validate the effectiveness on 5 types of downstream segmentation tasks including natural clean/corrupted images, medical images, camouflaged images and robotic images. Our proposed method is task-agnostic in nature and outperforms pre-trained SAM and state-of-the-art domain adaptation methods on almost all downstream tasks with the same testing prompt inputs.

Vision Transformer (ViT) has emerged as a prominent architecture for various computer vision tasks. In ViT, we divide the input image into patch tokens and process them through a stack of self-attention blocks. However, unlike Convolutional Neural Network (CNN), ViT’s simple architecture has no informative inductive bias (e.g., locality, etc.). Due to this, ViT requires a large amount of data for pre-training. Various data-efficient approaches (DeiT) have been proposed to train ViT on balanced datasets effectively. However, limited literature discusses the use of ViT for datasets with long-tailed imbalances. In this work, we introduce DeiT-LT to tackle the problem of training ViTs from scratch on long-tailed datasets. In DeiT-LT, we introduce an efficient and effective way of distillation from CNN via distillation \texttt{DIST} token by using out-of-distribution images and re-weighting the distillation loss to enhance focus on tail classes. This leads to the learning of local CNN-like features in early ViT blocks, improving generalization for tail classes. Further, to mitigate overfitting, we propose distilling from a flat CNN teacher, which leads to learning low-rank generalizable features for DIST tokens across all ViT blocks. With the proposed DeiT-LT scheme, the distillation DIST token becomes an expert on the tail classes, and the classifier CLS token becomes an expert on the head classes. The experts help to effectively learn features corresponding to both the majority and minority classes using a distinct set of tokens within the same ViT architecture. We show the effectiveness of DeiT-LT for training ViT from scratch on datasets ranging from small-scale CIFAR-10 LT to large-scale iNaturalist-2018. Project Page: https://rangwani-harsh.github.io/DeiT-LT.

This paper introduces Unified Language-driven Zero-shot Domain Adaptation (ULDA), a novel task setting that enables a single model to adapt to diverse target domains without explicit domain-ID knowledge. We identify the constraints in the existing language-driven zero-shot domain adaptation task, particularly the requirement for domain IDs and domain-specific models, which may restrict flexibility and scalability. To overcome these issues, we propose a new framework for ULDA, consisting of Hierarchical Context Alignment (HCA), Domain Consistent Representation Learning (DCRL), and Text-Driven Rectifier (TDR). These components work synergistically to align simulated features with target text across multiple visual levels, retain semantic correlations between different regional representations, and rectify biases between simulated and real target visual features, respectively. Our extensive empirical evaluations demonstrate that this framework achieves competitive performance in both settings, surpassing even the model that requires domain-ID, showcasing its superiority and generalization ability. The proposed method is not only effective but also maintains practicality and efficiency, as it does not introduce additional computational costs during inference. The code and models will be publicly available.

We study source-free unsupervised domain adaptation (SFUDA) for semantic segmentation, which aims to adapt a source-trained model to the target domain without accessing the source data. Many works have been proposed to address this challenging problem, among which uncertainty based self-training is a predominant approach. However, without comprehensive denoising mechanisms, they still largely fall into biased estimates when dealing with different domains and confirmation bias. In this paper, we observe that pseudo-label noise is mainly contained in unstable samples in which the predictions of most pixels undergo significant variations during self-training. Inspired by this, we propose a novel mechanism to denoise unstable samples with stable ones. Specifically, we introduce the Stable Neighbor Denoising (SND) approach, which effectively discovers highly correlated stable and unstable samples by nearest neighbor retrieval and guides the reliable optimization of unstable samples by bi-level learning. Moreover, we compensate for the stable set by object-level object paste, which can further eliminate the bias caused by less learned classes. Our SND enjoys two advantages. First, SND does not require a specific segmentor structure, endowing its universality. Second, SND simultaneously addresses the issues of class, domain, and confirmation biases during adaptation, ensuring its effectiveness. Extensive experiments show that SND consistently outperforms state-of-the-art methods in various SFUDA semantic segmentation settings. In addition, SND can be easily integrated with other approaches, obtaining further improvements. The source code will be publicly available.

Generalization to new domains not seen during training is one of the long-standing challenges in deploying neural networks in real-world applications. Existing generalization techniques either necessitate external images for augmentation, and/or aim at learning invariant representations by imposing various alignment constraints. Large-scale pretraining has recently shown promising generalization capabilities, along with the potential of binding different modalities. For instance, the advent of vision-language models like CLIP has opened the doorway for vision models to exploit the textual modality. In this paper, we introduce a simple framework for generalizing semantic segmentation networks by employing language as the source of randomization. Our recipe comprises three key ingredients: (i) the preservation of the intrinsic CLIP robustness through minimal fine-tuning, (ii) language-driven local style augmentation, and (iii) randomization by locally mixing the source and augmented styles during training. Extensive experiments report state-of-the-art results on various generalization benchmarks.

Prompt tuning represents a valuable technique for adapting pre-trained visual-language models (VLM) to various downstream tasks.Recent advancements in CoOp-based methods propose a set of learnable domain-shared or image-conditional textual tokens to facilitate the generation of task-specific textual classifiers. However, those textual tokens have a limited generalization ability regarding unseen domains, as they cannot dynamically adjust to the distribution of testing classes.To tackle this issue, we present a novel Textual-based Class-aware Prompt tuning(TCP) that explicitly incorporates prior knowledge about classes to enhance their discriminability.The critical concept of TCP involves leveraging Textual Knowledge Embedding (TKE) to map the high generalizability of class-level textual knowledge into class-aware textual tokens.By seamlessly integrating these class-aware prompts into the Text Encoder, a dynamic class-aware classifier is generated to enhance discriminability for unseen domains.During inference, TKE dynamically generates class-aware prompts related to the unseen classes.Comprehensive evaluations demonstrate that TKE serves as a plug-and-play module effortlessly combinable with existing methods. Furthermore, TCP consistently achieves superior performance while demanding less training time.

Adapters provide an efficient and lightweight mechanism for adapting trained transformer models to a variety of different tasks. However, they have often been found to be outperformed by other adaptation mechanisms including low-rank adaptation. In this paper, we provide an in-depth study of adapters, their internal structure, as well as various implementation choices. We uncover pitfalls for using adapters and suggest a concrete, improved adapter architecture, called Adapter+, that not only outperforms previous adapter implementations but surpasses a number of other, more complex adaptation mechanisms in several challenging settings. Despite this, our suggested adapter is highly robust and, unlike previous work, requires little to no manual intervention when addressing a novel scenario. Adapter+ reaches state-of-the-art average accuracy on the VTAB benchmark, even without a per-task hyperparameter optimization.

Online Continual Learning (CL) solves the problem of learning the ever-emerging new classification tasks from a continuous data stream. Unlike its offline counterpart, in online CL, the training data can only be seen once. Most existing online CL research regards catastrophic forgetting (i.e., model stability) as almost the only challenge. In this paper, we argue that the model's capability to acquire new knowledge (i.e., model plasticity) is another challenge in online CL. While replay-based strategies have been shown to be effective in alleviating catastrophic forgetting, there is a notable gap in research attention toward improving model plasticity. To this end, we propose Collaborative Continual Learning (CCL), a collaborative learning based strategy to improve the model's capability in acquiring new concepts. Additionally, we introduce Distillation Chain (DC), a collaborative learning scheme to boost the training of the models. We adapt CCL-DC to existing representative online CL works. Extensive experiments demonstrate that even if the learners are well-trained with state-of-the-art online CL methods, our strategy can still improve model plasticity dramatically, and thereby improve the overall performance by a large margin. The source code of our work is available at https://github.com/maorong-wang/CCL-DC.

The emergence of attention-based transformer models has led to their extensive use in various tasks, due to their superior generalization and transfer properties. Recent research has demonstrated that such models, when prompted appropriately, are excellent for few-shot inference. However, such techniques are under-explored for dense prediction tasks like semantic segmentation. In this work, we examine the effectiveness of prompting a transformer-decoder with learned visual prompts for the generalized few-shot segmentation (GFSS) task. Our goal is to achieve strong performance not only on novel categories with limited examples, but also to retain performance on base categories. We propose an approach to learn visual prompts with limited examples. These learned visual prompts are used to prompt a multiscale transformer decoder to facilitate accurate dense predictions. Additionally, we introduce a unidirectional causal attention mechanism between the novel prompts, learned with limited examples, and the base prompts, learned with abundant data. This mechanism enriches the novel prompts without deteriorating the base class performance. Overall, this form of prompting helps us achieve state-of-the-art performance for GFSS on two different benchmark datasets: COCO-$20^i$ and Pascal-$5^i$, without the need for test-time optimization (or transduction). Furthermore, test-time optimization leveraging unlabelled test data can be used to improve the prompts, which we refer to as transductive prompt tuning.

While fine-tuning is a de facto standard method for training deep neural networks, it still suffers from overfitting when using small target datasets. Previous methods improve fine-tuning performance by maintaining knowledge of the source datasets or introducing regularization terms such as contrastive loss. However, these methods require auxiliary source information (e.g., source labels or datasets) or heavy additional computations. In this paper, we propose a simple method called adaptive random feature regularization (AdaRand). AdaRand helps the feature extractors of training models to adaptively change the distribution of feature vectors for downstream classification tasks without auxiliary source information and with reasonable computation costs. To this end, AdaRand minimizes the gap between feature vectors and random reference vectors that are sampled from class-conditional Gaussian distributions. Furthermore, AdaRand dynamically updates the conditional distribution to follow the currently updated feature extractors and balance the distance between classes in feature spaces. Our experiments show that AdaRand outperforms the other fine-tuning regularization requiring auxiliary source information and heavy computation costs.

In this paper, we tackle the task of category-agnostic pose estimation (CAPE), which aims to predict poses for objects of any category with few annotated samples. Previous works either rely on local matching between features of support and query samples or require support keypoint identifier. The former is prone to overfitting due to its sensitivity to sparse samples, while the latter is impractical for the open-world nature of the task. To overcome these limitations, we propose ESCAPE - a Bayesian framework that learns a prior over the features of keypoints. The prior can be expressed as a mixture of super-keypoints, each being a high-level abstract keypoint that captures the statistics of semantically related keypoints from different categories. We estimate the super-keypoints from base categories and use them in adaptation to novel categories. The adaptation to an unseen category involves two steps: first, we match each novel keypoint to a related super-keypoint; and second, we transfer the knowledge encoded in the matched super-keypoints to the novel keypoints. For the first step, we propose a learnable matching network to capture the relationship between the novel keypoints and the super-keypoints, resulting in a more reliable matching. ESCAPE mitigates overfitting by directly transferring learned knowledge to novel categories while it does not use keypoint identifiers.

Domain Generalization (DG) aims to resolve distribution shifts between source and target domains, and current DG methods are default to the setting that data from source and target domains share identical categories. Nevertheless, there exists unseen classes from target domains in practical scenarios. To address this issue, Open Set Domain Generalization (OSDG) has emerged and several methods have been exclusively proposed. However, most existing methods adopt complex architectures with slight improvement compared with DG methods. Recently, vision-language models (VLMs) have been introduced in DG following the fine-tuning paradigm, but consume huge training overhead with large vision models. Therefore, in this paper, we innovate to transfer knowledge from VLMs to lightweight vision models and improve the robustness by introducing Perturbation Distillation (PD) from three perspectives, including Score, Class and Instance (SCI), named SCI-PD. Moreover, previous methods are oriented by the benchmarks with identical and fixed splits, ignoring the divergence between source domains. These methods are revealed to suffer from sharp performance decay with our proposed new benchmark Hybrid Domain Generalization (HDG) and a novel metric $H^{2}$-CV, which construct various splits to comprehensively assess the robustness of algorithms. Extensive experiments demonstrate that our method outperforms state-of-the-art algorithms on multiple datasets, especially improving the robustness when confronting data scarcity.

Multi-domain generalization (mDG) is universally aimed to minimize the discrepancy between training and testing distributions to enhance marginal-to-label distribution mapping. However, existing mDG literature lacks a general learning objective paradigm and often imposes constraints on static target marginal distributions. In this paper, we propose to leverage a $Y$-mapping $\psi$ to relax the constraint. We rethink the learning objective for mDG and design a new **general learning objective** to interpret and analyze most existing mDG wisdom. This general objective is bifurcated into two synergistic amis: learning domain-independent conditional features and maximizing a posterior. Explorations also extend to two effective regularization terms that incorporate prior information and suppress invalid causality, alleviating the issues that come with relaxed constraints. We theoretically contribute an upper bound for the domain alignment of domain-independent conditional features, disclosing that many previous mDG endeavors actually **optimize partially the objective** and thus lead to limited performance. As such, our study distills a general learning objective into four practical components, providing a general, robust, and flexible mechanism to handle complex domain shifts. Extensive empirical results indicate that the proposed objective with $Y$-mapping leads to substantially better mDG performance in various downstream tasks, including regression, segmentation, and classification.

Deep neural networks have shown great success in representation learning. However, when learning with noisy labels (LNL), they can easily overfit and fail to generalize to new data. This paper introduces a simple and effective method, named Learning to Bootstrap (L2B), which enables models to bootstrap themselves using their own predictions without being adversely affected by erroneous pseudo-labels. It achieves this by dynamically adjusting the importance weight between real observed and generated labels, as well as between different samples through meta-learning. Unlike existing instance reweighting methods, the key to our method lies in a new, versatile objective that enables implicit relabeling concurrently, leading to significant improvements without incurring additional costs. L2B offers several benefits over the baseline methods. It yields more robust models that are less susceptible to the impact of noisy labels by guiding the bootstrapping procedure more effectively. It better exploits the valuable information contained in corrupted instances by adapting the weights of both instances and labels. Furthermore, L2B is compatible with existing LNL methods and delivers competitive results spanning natural and medical imaging tasks including classification and segmentation under both synthetic and real-world noise. Extensive experiments demonstrate that our method effectively mitigates the challenges of noisy labels, often necessitating few to no validation samples, and is well generalized to other tasks such as image segmentation. This not only positions it as a robust complement to existing LNL techniques but also underscores its practical applicability. The code and models are available at https://github.com/yuyinzhou/l2b.

Category-agnostic pose estimation (CAPE) aims to predict keypoints for arbitrary classes given a few support images annotated with keypoints. Existing methods only rely on the features extracted at support keypoints to predict or refine the keypoints on query image, but a few support feature vectors are local and inadequate for CAPE. Considering that human can quickly perceive potential keypoints of arbitrary objects, we propose a novel framework for CAPE based on such potential keypoints (named as meta-points). Specifically, we maintain learnable embeddings to capture inherent information of various keypoints, which interact with image feature maps to produce meta-points without any support. The produced meta-points could serve as meaningful potential keypoints for CAPE. Due to the inevitable gap between inherency and annotation, we finally utilize the identities and details offered by support keypoints to assign and refine meta-points to desired keypoints in query image. In addition, we propose a progressive deformable point decoder and a slacked regression loss for better prediction and supervision. Our novel framework not only reveals the inherency of keypoints but also outperforms existing methods of CAPE. Comprehensive experiments and in-depth studies on large-scale MP-100 dataset demonstrate the effectiveness of our framework.

Deep Neural Networks (DNNs) have become pivotal in various fields, especially in computer vision, outperforming previous methodologies. A critical challenge in their deployment is the bias inherent in data across different domains, such as image style, and environmental conditions, leading to domain gaps. This necessitates techniques for learning general representations from biased training data, known as domain generalization. This paper presents Attend to eXpert Prompts (A2XP), a novel approach for domain generalization that preserves the privacy and integrity of the network architecture. A2XP consists of two phases: Expert Adaptation and Domain Generalization. In the first phase, prompts for each source domain are optimized to guide the model towards the optimal direction. In the second phase, two embedder networks are trained to effectively amalgamate these expert prompts, aiming for an optimal output. Our extensive experiments demonstrate that A2XP achieves state-of-the-art results over existing non-private domain generalization methods.The experimental results validate that the proposed approach not only tackles the domain generalization challenge in DNNs but also offers a privacy-preserving, efficient solution to the broader field of computer vision.

Class-Incremental Learning (CIL) requires a learning system to continually learn new classes without forgetting. Despite the strong performance of Pre-Trained Models (PTMs) in CIL, a critical issue persists: learning new classes often results in the overwriting of old ones. Excessive modification of the network causes forgetting, while minimal adjustments lead to an inadequate fit for new classes. As a result, it is desired to figure out a way of efficient model updating without harming former knowledge.In this paper, we propose ExpAndable Subspace Ensemble (EASE) for PTM-based CIL. To enable model updating without conflict, we train a distinct lightweight adapter module for each new task, aiming to create task-specific subspaces. These adapters span a high-dimensional feature space, enabling joint decision-making across multiple subspaces. As data evolves, the expanding subspaces render the old class classifiers incompatible with new-stage spaces.Correspondingly, we design a semantic-guided prototype complement strategy that synthesizes old classes' new features without using any old class instance. Extensive experiments on seven benchmark datasets verify EASE's state-of-the-art performance.

In this paper, we propose a novel Visual Reference Prompt (VRP) encoder that empowers the Segment Anything Model (SAM) to utilize annotated reference images as prompts for segmentation, creating the VRP-SAM model. In essence, VRP-SAM can utilize annotated reference images to comprehend specific objects and perform segmentation of specific objects in target image. It is note that the VRP encoder can support a variety of annotation formats for reference images, including \textbf{point}, \textbf{box}, \textbf{scribble}, and \textbf{coarse mask}. VRP-SAM achieves a breakthrough within the SAM framework by extending its versatility and applicability while preserving SAM's inherent strengths, thus enhancing user-friendliness. To enhance the generalization ability of VRP-SAM, the VRP encoder adopts a meta-learning strategy. To validate the effectiveness of VRP-SAM, we conducted extensive empirical studies on the Pascal and COCO datasets. Remarkably, VRP-SAM achieved state-of-the-art performance in visual reference segmentation with minimal learnable parameters. Furthermore, VRP-SAM demonstrates strong generalization capabilities, allowing it to perform segmentation of unseen objects and enabling cross-domain segmentation. The source code and models will be available at https://github.com/syp2ysy/VRP-SAM

Cross-domain few-shot learning (CDFSL) aims to acquire knowledge from limited training data in the target domain by leveraging prior knowledge transferred from source domains with abundant training samples. CDFSL faces challenges in transferring knowledge across dissimilar domains and fine-tuning models with limited training data. To address these challenges, we initially extend the analysis of loss landscapes from the parameter space to the representation space, which allows us to simultaneously interpret the transferring and fine-tuning difficulties of CDFSL models. We observe that sharp minima in the loss landscapes of the representation space result in representations that are hard to transfer and fine-tune. Moreover, existing flatness-based methods have limited generalization ability due to their short-range flatness. To enhance the transferability and facilitate fine-tuning, we introduce a simple yet effective approach to achieve long-range flattening of the minima in the loss landscape. This approach considers representations that are differently normalized as minima in the loss landscape and flattens the high-loss region in the middle by randomly sampling interpolated representations. We implement this method as a new normalization layer that replaces the original one in both CNNs and ViTs. This layer is simple and lightweight, introducing only a minimal number of additional parameters. Experimental results on 8 datasets demonstrate that our approach outperforms state-of-the-art methods in terms of average accuracy. Moreover, our method achieves performance improvements of up to 9\% compared to the current best approaches on individual datasets. Our code will be released.

Deep learning has achieved remarkable progress in various applications, heightening the importance of safeguarding the intellectual property (IP) of well-trained models. It entails not only authorizing usage but also ensuring the deployment of models in authorized data domains, i.e., making models exclusive to certain target domains. Previous methods necessitate concurrent access to source training data and target unauthorized data when performing IP protection, making them risky and inefficient for decentralized private data. In this paper, we target a practical setting where only a well-trained source model is available and investigate how we can realize IP protection. To achieve this, we propose a novel MAsk Pruning (MAP) framework. MAP stems from an intuitive hypothesis, i.e., there are target-related parameters in a well-trained model, locating and pruning them is the key to IP protection. Technically, MAP freezes the source model and learns a target-specific binary mask to prevent unauthorized data usage while minimizing performance degradation on authorized data. Moreover, we introduce a new metric aimed at achieving a better balance between source and target performance degradation. To verify the effectiveness and versatility, we have evaluated MAP in a variety of scenarios, including vanilla source-available, practical source-free, and challenging data-free. Extensive experiments indicate that MAP yields new state-of-the-art performance.

Domain Generalization (DG) aims to develop a versatile model capable of performing well on unseen target domains. Recent advancements in pre-trained Visual Foundation Models (VFMs), such as CLIP, show significant potential in enhancing the generalization abilities of deep models. Although there is a growing focus on VFM-based domain prompt tuning for DG, effectively learning prompts that disentangle invariant features across all domains remains a major challenge. In this paper, we propose addressing this challenge by leveraging the controllable and flexible language prompt of the VFM. Observing that the text modality of VFMs is inherently easier to disentangle, we introduce a novel text feature guided visual prompt tuning framework. This framework first automatically disentangles the text prompt using a large language model (LLM) and then learns domain-invariant visual representation guided by the disentangled text feature. Moreover, we also devise domain-specific prototype learning to fully exploit domain-specific information to combine with the invariant feature prediction. Extensive experiments on mainstream DG datasets, namely PACS, VLCS, OfficeHome, DomainNet and TerraInc, demonstrate that the proposed method achieves superior performances to state-of-the-art DG methods. Our source code is available in the supplementary materials.

Few-shot segmentation performance declines substantially when facing images from a domain different than the training domain, effectively limiting real-world use cases. To alleviate this, recently cross-domain few-shot segmentation (CD-FSS) has emerged. Works that address this task mainly attempted to learn segmentation on a source domain in a manner that generalizes across domains. Surprisingly, we can outperform these approaches while eliminating the training stage and removing their main segmentation network. We show test-time task-adaption is the key for successful CD-FSS instead. Task-adaption is achieved by appending small networks to the feature pyramid of a conventionally classification-pretrained backbone. To avoid overfitting to the few labeled samples in supervised fine-tuning, consistency across augmented views of input images serves as guidance while learning the parameters of the attached layers. Despite our self-restriction not to use any images other than the few labeled samples at test time, we achieve new state-of-the-art performance in CD-FSS, evidencing the need to rethink approaches for the task.

Continual Learning (CL) enables machine learning models to learn from continuously shifting new training data in absence of data from old tasks. Recently, pre-trained vision transformers combined with prompt tuning have shown promise for overcoming catastrophic forgetting in CL. These approaches rely on a pool of learnable prompts which can be inefficient in sharing knowledge across tasks leading to inferior performance. In addition, the lack of fine-grained layer specific prompts does not allow these to fully express the strength of the prompts for CL. We address these limitations by proposing ConvPrompt, a novel convolutional prompt creation mechanism that maintains layer-wise shared embeddings, enabling both layer-specific learning and better concept transfer across tasks. The intelligent use of convolution enables us to maintain a low parameter overhead without compromising performance. We further leverage Large Language Models to generate fine-grained text descriptions of each category which are used to get task similarity and dynamically decide the number of prompts to be learned. Extensive experiments demonstrate the superiority of ConvPrompt and improves SOTA by $\sim 3$% with significantly less parameter overhead. We also perform strong ablation over various modules to disentangle the importance of different components. Our code will be made public.

Generative Zero-Shot Learning (ZSL) learns a generator to synthesize visual samples for unseen classes, which is an effective way to advance ZSL. However, existing generative methods rely on the conditions of Gaussian noise and the predefined semantic prototype, which limit the generator only optimized on specific seen classes rather than characterizing each visual instance, resulting in poor generalizations (e.g., overfitting to seen classes). To address this issue, we propose a novel Visual-Augmented Dynamic Semantic prototype method (termed VADS) to boost the generator to learn accurate semantic-visual mapping by fully exploiting the visual-augmented knowledge into semantic conditions. In detail, VADS consists of two modules: (1) Visual-aware Domain Knowledge Learning module (VDKL) learns the local bias and global prior of the visual features (referred to as domain visual knowledge), which replace pure Gaussian noise to provide richer prior noise information; (2) Vision-Oriented Semantic Updation module (VOSU) updates the semantic prototype according to the visual representations of the samples. Ultimately, we concatenate their output as a dynamic semantic prototype, which serves as the condition of the generator. Extensive experiments demonstrate that our VADS achieves superior CZSL and GZSL performances on three prominent datasets and outperforms other state-of-the-art methods with averaging increases by 6.4\%, 5.9\% and 4.2\% on SUN, CUB and AWA2, respectively.

Continual learning requires the model to learn multiple tasks sequentially. In continual learning, the model should possess the ability to maintain its performance on old tasks (stability) and the ability to adapt to new tasks continuously (plasticity). Recently, parameter-efficient fine-tuning (PEFT), which involves freezing a pre-trained model and injecting a small number of learnable parameters to adapt to downstream tasks, has gained increasing popularity in continual learning. Although existing continual learning methods based on PEFT have demonstrated superior performance compared to those not based on PEFT, most of them do not consider how to eliminate the interference of the new task on the old tasks, which inhibits the model from making a good trade-off between stability and plasticity. In this work, we propose a new PEFT method, called interference-free low-rank adaptation (InfLoRA), for continual learning. InfLoRA injects a small number of parameters to reparameterize the pre-trained weights and shows that fine-tuning these injected parameters is equivalent to fine-tuning the pre-trained weights within a subspace. Furthermore, InfLoRA designs this subspace to eliminate the interference of the new task on the old tasks, making a good trade-off between stability and plasticity. Experimental results show that InfLoRA outperforms existing state-of-the-art continual learning methods on multiple datasets.

Source-free domain adaptation (SFDA) assumes that model adaptation only accesses the well-learned source model and unlabeled target instances for knowledge transfer. However, cross-domain distribution shift easily triggers invalid discriminative semantics from source model on recognizing the target samples. Hence, understanding the specific content of discriminative pattern and adjusting their representation in target domain become the important key to overcome SFDA. To achieve such a vision, this paper proposes a novel explanation paradigm ''Discriminative Pattern Calibration (DPC)'' mechanism on solving SFDA issue. Concretely, DPC first utilizes learning network to infer the discriminative regions on the target images and specifically emphasizes them in feature space to enhance their representation. Moreover, DPC relies on the attention-reversed mixup mechanism to augment more samples and improve the robustness of the classifier. Considerable experimental results and studies suggest that the effectiveness of our DPC in enhancing the performance of existing SFDA baselines.

Deep neural networks (DNNs) struggle to learn in dynamic settings because they mainly rely on static datasets. Continual learning (CL) aims to overcome this limitation by enabling DNNs to incrementally accumulate knowledge. A widely adopted scenario in CL is class-incremental learning (CIL), where DNNs are required to sequentially learn more classes. Among the various strategies in CL, replay methods, which revisit previous classes, stand out as the only effective ones in CIL. Other strategies, such as architectural modifications to segregate information across weights and protect them from change, are ineffective in CIL. This is because they need additional information during testing to select the correct network parts to use. In this paper, we propose NICE, Neurogenesis Inspired Contextual Encoding, a replay-free architectural method inspired by adult neurogenesis in the hippocampus. NICE groups neurons in the DNN based on different maturation stages and infers which neurons to use during testing without any additional signal. Through extensive experiments across 6 datasets and 3 architectures, we show that NICE performs on par with or often outperforms replay methods. We also make the case that neurons exhibit highly distinctive activation patterns for the classes in which they specialize, enabling us to determine when they should be used. The code is available at https://github.com/BurakGurbuz97/NICE.

To accommodate real-world dynamics, artificial intelligence systems need to cope with sequentially arriving content in an online manner. Beyond regular Continual Learning (CL) attempting to address catastrophic forgetting with offline training of each task, Online Continual Learning (OCL) is a more challenging yet realistic setting that performs CL in a one-pass data stream. Current OCL methods primarily rely on memory replay of old training samples. However, a notable gap from CL to OCL stems from the additional overfitting-underfitting dilemma associated with the use of rehearsal buffers: the inadequate learning of new training samples (underfitting) and the repeated learning of a few old training samples (overfitting). To this end, we introduce a novel approach, Multi-level Online Sequential Experts (MOSE), which cultivates the model as stacked sub-experts, integrating multi-level supervision and reverse self-distillation. Supervision signals across multiple stages facilitate appropriate convergence of the new task while gathering various strengths from experts by knowledge distillation mitigates the performance decline of old tasks. MOSE demonstrates remarkable efficacy in learning new samples and preserving past knowledge through multi-level experts, thereby significantly advancing OCL performance over state-of-the-art baselines (e.g., up to 7.3% on Split CIFAR-100 and 6.1% on Split Tiny-ImageNet).

Efficient transfer learning (ETL) is receiving increasing attention to adapt large pre-trained language-vision models on downstream tasks with a few labeled samples. While significant progress has been made, we reveal that state-of-the-art ETL approaches exhibit strong performance only in narrowly-defined experimental setups, and with a careful adjustment of hyperparameters based on a large corpus of labeled samples. In particular, we make two interesting, and surprising empirical observations. First, to outperform a simple Linear Probing baseline, these methods require to optimize their hyper-parameters on each target task. And second, they typically underperform --sometimes dramatically-- standard zero-shot predictions in the presence of distributional drifts. Motivated by the unrealistic assumptions made in the existing literature, i.e., access to a large validation set and case-specific grid-search for optimal hyperparameters, we propose a novel approach that meets the requirements of real-world scenarios. More concretely, we introduce a CLass-Adaptive linear Probe (CLAP) objective, whose balancing term is optimized via an adaptation of the general Augmented Lagrangian method tailored to this context. We comprehensively evaluate CLAP on a broad span of datasets and scenarios, demonstrating that it consistently outperforms SoTA approaches, while yet being a much more efficient alternative.

We approach the challenge of addressing semi-supervised domain generalization (SSDG). Specifically, our aim is to obtain a model that learns domain-generalizable features by leveraging a limited subset of labelled data alongside a substantially larger pool of unlabeled data. Existing domain generalization (DG) methods which are unable to exploit unlabelled data perform poorly compared to semi-supervised learning (SSL) methods under SSDG setting. Nevertheless, SSL methods have a considerable room for performance improvement when compared to fully-supervised DG training. To tackle this underexplored, yet highly practical problem of SSDG, we make the following core contributions. First, we propose a feature-based conformity technique that matches the posterior distributions from the feature space with the pseudo-label from the model's output space. Second, we develop a semantics alignment loss to learn semantically-compatible representations by regularizing the semantic structure in the feature space. Our method is plug-and-play and can be readily integrated with different SSL-based SSDG baselines without introducing any additional parameters. Extensive experimental results across five challenging DG benchmarks with four strong SSL baselines suggest that our method provides consistent and notable gains in two different SSDG settings. Our code will be made publicly available.

Test-time adaptation (TTA) is a technique to improve the performance of a pre-trained source model on a target distribution without using any labeled data. However, existing self-trained TTA methods often face the challenges of unreliable pseudo-labels and unstable model optimization. In this paper, we propose an Improved Self-Training (IST) approach, which addresses these challenges by enhancing the pseudo-label quality and stabilizing the adaptation process. Specifically, we use a simple augmentation strategy to generate multiple views of each test sample, and construct a graph structure to correct the pseudo-labels based on the similarity of the latent features. Moreover, we adopt a parameter moving average scheme to smooth the model updates and prevent catastrophic forgetting. Instead of using a model with fixed label space, we explore the adaptability of the foundation model CLIP to various downstream tasks at test time. Extensive experiments on various benchmarks show that IST can achieve significant and consistent improvements over the existing TTA methods in classification, detection, and segmentation tasks.

Source-Free Domain Adaptation (SFDA) aims to adapt a source model for a target domain, with only access to unlabeled target training data and the source model pre-trained on a supervised source domain. Relying on pseudo labeling and/or auxiliary supervision, conventional methods are inevitably error-prone. To mitigate this limitation, in this work we for the first time explore the potentials of off-the-shelf vision-language (ViL) multimodal models (e.g., CLIP) with rich whilst heterogeneous knowledge. We find that directly applying the ViL model to the target domain in a zero-shot fashion is unsatisfactory, as it is not specialized for this particular task but largely generic. To make it task specific, we propose a novel Distilling multImodal Foundation mOdel (DIFO) approach. Specifically, DIFO alternates between two steps during adaptation: (i) Customizing the ViL model by maximizing the mutual information with the target model in a prompt learning manner, (ii) Distilling the knowledge of this customized ViL model to the target model. For more fine-grained and reliable distillation, we further introduce two effective regularization terms, namely most likely category encouragement and predictive consistency. Extensive experiments show that DIFO significantly outperforms the state-of-the-art alternatives. Our source code will be released.

Regression tasks in computer vision, such as age estimation or counting, are often formulated into classification by quantizing the target space into classes. Yet real-world data is often imbalanced -- the majority of training samples lie in a head range of target values, while a minority of samples span a usually larger tail range. By selecting the class quantization, one can adjust imbalanced regression targets into balanced classification outputs, though there are trade-offs in balancing classification accuracy and quantization error. To improve regression performance over the entire range of data, we propose to construct hierarchical classifiers for solving imbalanced regression tasks. The fine-grained classifiers limit the quantization error while being modulated by the coarse predictions to ensure high accuracy. Standard hierarchical classification approaches, however, when applied to the regression problem, fail to ensure that predicted ranges remain consistent across the hierarchy. As such, we propose a range-preserving distillation process that can effectively learn a single classifier from the set of hierarchical classifiers. Our novel hierarchical classification adjustment (HCA) for imbalanced regression shows superior results on three diverse tasks: age estimation, crowd counting and depth estimation. Code is available at https://github.com/xhp-hust-2018-2011/HCA.

Continual test-time domain adaptation (CTTA) aims to adapt the source pre-trained model to a continually changing target domain without additional data acquisition or labeling costs. This issue necessitates an initial performance enhancement within the present domain without labels while concurrently averting an excessive bias toward the current domain. Such bias exacerbates catastrophic forgetting and diminishes the generalization ability to future domains. To tackle the problem, this paper designs a versatile framework to capture high-quality supervision signals from three aspects: 1) The adaptive thresholds are employed to determine the reliability of pseudo-labels; 2) The knowledge from the source pre-trained model is utilized to adjust the unreliable one, and 3) By evaluating past supervision signals, we calculate a diversity score to ensure subsequent generalization. In this way, we form a complete supervisory signal generation framework, which can capture the current domain discriminative and reserve generalization in future domains. Finally, to avoid catastrophic forgetting, we design a weighted soft parameter alignment method to explore the knowledge from the source model. Extensive experimental results demonstrate that our method performs well on several benchmark datasets.

In this paper, we focus on a challenging Online Task-Free Class Incremental Learning (OTFCIL) problem. Different from the existing methods that continuously learn the feature space from data streams, we propose a novel compute-and-align paradigm for the OTFCIL. It first computes an optimal geometry, i.e., the class prototype distribution, for classifying existing classes and updates it when new classes emerge, and then trains a DNN model by aligning its feature space to the optimal geometry. To this end, we develop a novel Dynamic Neural Collapse (DNC) algorithm to compute and update the optimal geometry. The DNC expands the geometry when new classes emerge without loss of the geometry optimality and guarantees the drift distance of old class prototypes with an explicit upper bound. Then, we propose a novel Dynamic feature space Self-Organization (DYSON) method containing three major components, including 1) a feature extractor, 2) a Dynamic Feature-Geometry Alignment (DFGA) module aligning the feature space to the optimal geometry computed by DNC and 3) a training-free class-incremental classifier derived from the DNC geometry. Experimental comparison results on four benchmark datasets, including CIFAR10, CIFAR100, CUB200, and CoRe50, demonstrate the efficiency and superiority of the DYSON method. The source code is provided in the supplementary material.

Out-of-Distribution (OOD) detection aims to address the excessive confidence prediction by neural networks by triggering an alert when the input sample deviates significantly from the training distribution (in-distribution), indicating that the output may not be reliable.Current OOD detection approaches explore all kinds of cues to identify OOD data, such as finding irregular patterns in the feature space, logit space, gradient space, or the raw image space. Surprisingly, we observe a linear trend between the OOD score produced by current OOD detection algorithms and the network features on several datasets.We conduct a thorough investigation, theoretically and empirically, to analyze and understand the meaning of such a linear trend in OOD detection.This paper proposes a Robust Test-time Linear method (RTL) to utilize such linear trends like a `free lunch' when we have a batch of data to perform OOD detection.By using a simple linear regression as a test time adaptation, we can make a more precise OOD prediction.We further propose an online variant of the proposed method, which achieves promising performance and is more practical for real applications. Theoretical analysis is given to prove the effectiveness of our methods.Extensive experiments on several OOD datasets show the efficacy of RTL for OOD detection tasks, significantly improving the results of base OOD detectors. Project will be available at https://github.com/kfan21/RTL.

Long-tail recognition is challenging because it requires the model to learn good representations from tail categories and address imbalances across all categories. In this paper, we propose a novel generative and fine-tuning framework, LTGC, to handle long-tail recognition via leveraging generated content. Firstly, inspired by the rich implicit knowledge in large-scale models (e.g., large language models, LLMs), LTGC leverages the power of these models to parse and reason over the original tail data to produce diverse tail-class content. We then propose several novel designs for LTGC to ensure the quality of the generated data and to efficiently fine-tune the model using both the generated and original data. The visualization demonstrates the effectiveness of the generation module in LTGC, which produces accurate and diverse tail data. Additionally, the experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tailed benchmarks.

Few-shot semantic segmentation (FSS) endeavors to segment unseen classes with only a few labeled samples. Current FSS methods are commonly built on the assumption that their training and application scenarios share similar domains, and their performances degrade significantly while applied to a distinct domain. To this end, we propose to leverage the cutting-edge foundation model, the Segment Anything Model (SAM), for generalization enhancement. The SAM however performs unsatisfactorily on domains that are distinct from its training data, which primarily comprise natural scene images, and it does not support automatic segmentation of specific semantics due to its interactive prompting mechanism. In our work, we introduce APSeg, a novel auto-prompt network for cross-domain few-shot semantic segmentation (CD-FSS), which is designed to be auto-prompted for guiding cross-domain segmentation. Specifically, we propose a Dual Prototype Anchor Transformation (DPAT) module that fuses pseudo query prototypes extracted based on cycle-consistency with support prototypes, allowing features to be transformed into a more stable domain-agnostic space. Additionally, a Meta Prompt (MPG) module is introduced to automatically generate prompt embeddings, eliminating the need for manual visual prompts. We build an efficient model which can be applied directly to target domains without fine-tuning. Extensive experiments on four cross-domain datasets show that our model outperforms the state-of-the-art CD-FSS method by 5.24\% and 3.10\% in average accuracy on 1-shot and 5-shot settings, respectively.

In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin LP++, step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yields, surprisingly, highly competitive few-shot CLIP performances. Furthermore, LP++ operates in black-box, relaxes intensive validation searches for the optimization hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP adaptation methods. Our code is available at: \url{https://github.com/FereshteShakeri/FewShot-CLIP-Strong-Baseline.git}.

The development of large vision-language models, notably CLIP, has catalyzed research into effective adaptation techniques, with a particular focus on soft prompt tuning. Conjointly, test-time augmentation, which utilizes multiple augmented views of a single image to enhance zero-shot generalization, is emerging as a significant area of interest. This has predominantly directed research efforts towards test-time prompt tuning. In contrast, we introduce a robust $\textbf{M}$eanShift for $\textbf{T}$est-time $\textbf{A}$ugmentation (MTA), which surpasses prompt-based methods without requiring this intensive training procedure. This positions MTA as an ideal solution for both standalone and API-based applications. Additionally, our method does not rely on ad hoc rules (e.g., confidence threshold) used in some previous test-time augmentation techniques to filter the augmented views. Instead, MTA incorporates a quality assessment variable for each view directly into its optimization process, termed as the inlierness score. This score is jointly optimized with a density mode seeking process, leading to an efficient training- and hyperparameter-free approach. We extensively benchmark our method on 15 datasets and demonstrate MTA's superiority and computational efficiency. Deployed easily as plug-and-play module on top of zero-shot models and state-of-the-art few-shot methods, MTA shows systematic and consistent improvements.

In this paper, we look at cross-domain few-shot classification which presents the challenging task of learning new classes in previously unseen domains with few labelled examples. Existing methods, though somewhat effective, encounter several limitations, which we alleviate through two significant improvements. First, we introduce a lightweight parameter-efficient adaptation strategy to address overfitting associated with fine-tuning a large number of parameters on small datasets. This strategy employs a linear transformation of pre-trained features, significantly reducing the trainable parameter count. Second, we replace the traditional nearest centroid classifier with a discriminative sample-aware loss function, enhancing the model's sensitivity to the inter- and intra-class variances within the training set for improved clustering in feature space. Empirical evaluations on the Meta-Dataset benchmark showcase that our approach not only improves accuracy up to 7.7\% and 5.3\% on previously seen and unseen datasets, respectively, but also achieves the above performance while being at least $\sim3\times$ more parameter-efficient than existing methods, establishing a new state-of-the-art in cross-domain few-shot learning. Our code is available at https://github.com/rashindrie/DIPA.

In order for reinforcement learning (RL) agents to be deployed in real-world environments, they must be able to generalize to unseen environments. However, RL struggles with out-of-distribution generalization, often due to over-fitting the particulars of the training environment. Although regularization techniques from supervised learning can be applied to avoid over-fitting, the differences between supervised learning and RL limit their application. To address this, we propose the Signal-to-Noise Ratio regulated Parameter Uncertainty Network (SNR PUN) for RL. We introduce SNR as a new measure of regularizing the parameter uncertainty of a network and provide a formal analysis explaining why SNR regularization works well for RL. We demonstrate the effectiveness of our proposed method to generalize in several simulated environments; and in a physical system showing the possibility of using SNR PUN for applying RL to real-world applications.

3D Object Detectors (3D-OD) are crucial for understanding the environment in many robotic tasks, especially autonomous driving. Including 3D information via Lidar sensors improves accuracy greatly. However, such detectors perform poorly on domains they were not trained on, i.e. different locations, sensors, weather, etc., limiting their reliability in safety-critical applications. There exist methods to adapt 3D-ODs to these domains; however, these methods treat 3D-ODs as a black box, neglecting underlying architectural decisions and source-domain training strategies. Instead, we dive deep into the details of 3D-ODs, focusing our efforts on fundamental factors that influence robustness prior to domain adaptation.We systematically investigate four design choices (and the interplay between them) often overlooked in 3D-OD robustness and domain adaptation: architecture, voxel encoding, data augmentations, and anchor strategies. We assess their impact on the robustness of nine state-of-the-art 3D-ODs across six benchmarks encompassing three types of domain gaps - sensor type, weather, and location.Our main findings are: (1) transformer backbones with local point features are more robust than 3D CNNs, (2) test-time anchor size adjustment is crucial for adaptation across geographical locations, significantly boosting scores without retraining, (3) source-domain augmentations allow the model to generalize to low-resolution sensors, and (4) surprisingly, robustness to bad weather is improved when training directly on more clean weather data than on training with bad weather data. We outline our main conclusions and findings to provide practical guidance on developing more robust 3D-ODs.

Pre-trained Vision-Language Models (VLMs) have served as excellent foundation models for transfer learning in diverse downstream tasks. However, tuning VLMs for few-shot generalization tasks faces a discrimination — generalization dilemma, i.e., general knowledge should be preserved and task-specific knowledge should be fine-tuned. How to precisely identify these two types of representations remains a challenge. In this paper, we propose a Multi-Modal Adapter (MMA) for VLMs to improve the alignment between representations from text and vision branches. MMA aggregates features from different branches into a shared feature space so that gradients can be communicated across branches. To determine how to incorporate MMA, we systematically analyze the discriminability and generalizability of features across diverse datasets in both the vision and language branches, and find that (1) higher layers contain discriminable dataset-specific knowledge, while lower layers contain more generalizable knowledge, and (2) language features are more discriminable than visual features, and there are large semantic gaps between the features of the two modalities, especially in the lower layers. Therefore, we only incorporate MMA to a few higher layers of transformers to achieve an optimal balance between discrimination and generalization. We evaluate the effectiveness of our approach on three tasks: generalization to novel classes, novel target datasets, and domain generalization. Compared to many state-of-the-art methods, our MMA achieves leading performance in all evaluations. Code is at https://github.com/ZjjConan/Multi-Modal-Adapter

Personalized Federated Learning (pFL) has emerged as a promising solution to tackle data heterogeneity across clients in FL. However, existing pFL methods either (1) introduce high computation and communication costs or (2) overfit to local data, which can be limited in scope and vulnerable to evolved test samples with natural distribution shifts. In this paper, we propose PerAda, a parameter-efficient pFL framework that reduces communication and computational costs and exhibits superior generalization performance, especially under test-time distribution shifts. PerAda reduces the costs by leveraging the power of pretrained models and only updates and communicates a small number of additional parameters from adapters. PerAda achieves high generalization by regularizing each client's personalized adapter with a global adapter, while the global adapter uses knowledge distillation to aggregate generalized information from all clients. Theoretically, we provide generalization bounds of PerAda, and we prove its convergence to stationary points under non-convex settings. Empirically, PerAda demonstrates higher personalized performance (+4.85\% on CheXpert) and enables better out-of-distribution generalization (+5.23\% on CIFAR-10-C) on different datasets across natural and medical domains compared with baselines, while only updating 12.6\% of parameters per model.

Low-shot image classification is a fundamental task in computer vision, and the emergence of large-scale vision-language models such as CLIP has greatly advanced the forefront of research in this field. However, most existing CLIP-based methods lack the flexibility to effectively incorporate other pre-trained models that encompass knowledge distinct from CLIP. To bridge the gap, this work proposes a simple and effective probabilistic model ensemble framework based on Gaussian processes, which have previously demonstrated remarkable efficacy in processing small data. We achieve the integration of prior knowledge by specifying the mean function with CLIP and the kernel function with an ensemble of deep kernels built upon various pre-trained models. By regressing the classification label directly, our framework enables analytical inference, straightforward uncertainty quantification, and principled hyper-parameter tuning. Through extensive experiments on standard benchmarks, we demonstrate that our method consistently outperforms competitive ensemble baselines regarding predictive performance. Additionally, we assess the robustness of our method and the quality of the yielded uncertainty estimates on out-of-distribution datasets. We also illustrate that our method, despite relying on label regression, still enjoys superior model calibration compared to most deterministic baselines.

Data-Free Knowledge Distillation (DFKD) has recently made remarkable advancements with its core principle of transferring knowledge from a teacher neural network to a student neural network without requiring access to the original data. Nonetheless, existing approaches encounter a significant challenge when attempting to generate samples from random noise inputs, which inherently lack meaningful information. Consequently, these models struggle to effectively map this noise to the ground-truth sample distribution, resulting in the production of low-quality data and imposing substantial time requirements for training the generator. In this paper, we propose a novel Noisy Layer Generation method (NAYER) which relocates the random source from the input to a noisy layer and utilizes the meaningful constant label-text embedding (LTE) as the input. {\color{black} LTE is generated by using the language model once, and then it is stored in memory for all subsequent training processes.} The significance of LTE lies in its ability to contain substantial meaningful inter-class information, enabling the generation of high-quality samples with only a few training steps. Simultaneously, the noisy layer plays a key role in addressing the issue of diversity in sample generation by preventing the model from overemphasizing the constrained label information. By reinitializing the noisy layer in each iteration, we aim to facilitate the generation of diverse samples while still retaining the method's efficiency, thanks to the ease of learning provided by LTE. Experiments carried out on multiple datasets demonstrate that our NAYER not only outperforms the state-of-the-art methods but also achieves speeds 5 to 15 times faster than previous approaches. The code is available at https://github.com/fw742211/nayer.

Federated Class-Incremental Learning (FCIL) is an underexplored yet pivotal issue, involving the dynamic addition of new classes in the context of federated learning. In this field, Data-Free Knowledge Transfer (DFKT) plays a crucial role in addressing catastrophic forgetting and data privacy problems. However, prior approaches lack the crucial synergy between DFKT and the model training phases, causing DFKT to encounter difficulties in generating high-quality data from a non-anchored latent space of the old task model. In this paper, we introduce LANDER (Label Text Centered Data-Free Knowledge Transfer) to address this issue by utilizing label text embeddings (LTE) produced by pretrained language models. Specifically, during the model training phase, our approach treats LTE as anchor points and constrains the feature embeddings of corresponding training samples around them, enriching the surrounding area with more meaningful information. In the DFKT phase, by using these LTE anchors, LANDER can synthesize more meaningful samples, thereby effectively addressing the forgetting problem. Additionally, instead of tightly constraining embeddings toward the anchor, the Bounding Loss is introduced to encourage sample embeddings to remain flexible within a defined radius. This approach preserves the natural differences in sample embeddings and mitigates the embedding overlap caused by heterogeneous federated settings. Extensive experiments conducted on CIFAR100, Tiny-ImageNet, and ImageNet demonstrate that LANDER significantly outperforms previous methods and achieves state-of-the-art performance in FCIL. The code is available at https://anonymous.4open.science/r/lander-fcil/.

Few-Shot Class Incremental Learning (FSCIL) is a task that requires a model to learn new classes incrementally without forgetting when only a few samples for each class are given. FSCIL encounters two significant challenges: catastrophic forgetting and overfitting, and these challenges have driven prior studies to primarily rely on shallow models, such as ResNet-18. Even though their limited capacity can mitigate both forgetting and overfitting issues, it leads to inadequate knowledge transfer during few-shot incremental sessions. In this paper, we argue that $\textit{large models such as vision and language transformers pre-trained}$ $\textit{on large datasets can be excellent few-shot incremental learners.}$ To this end, we propose a novel FSCIL framework called $\textbf{PriViLege}$, $\textbf{Pr}$e-tra$\textbf{i}$ned $\textbf{Vi}$sion and $\textbf{L}$anguage transformers with prompting functions and knowl$\textbf{e}$d$\textbf{ge}$ distillation. Our framework effectively addresses the challenges of catastrophic forgetting and overfitting in large models through new pre-trained knowledge tuning (PKT) and two losses: entropy-based divergence loss and semantic knowledge distillation loss.Experimental results show that the proposed PriViLege significantly outperforms the existing state-of-the-art methods with a large margin, \textit{e.g.}, $+9.38$% in CUB200, $+20.58$% in CIFAR-100, and $+13.36$% in miniImageNet. The code will be publicly released after the review.

Pseudo-label-based semi-supervised learning (SSL) algorithms trained on a class-imbalanced set face two cascading challenges: 1) Classifiers tend to be biased towards majority classes, and 2) Biased pseudo-labels are used for training. It is difficult to appropriately re-balance the classifiers in SSL because the class distribution of an unlabeled set is often unknown and could be mismatched with that of a labeled set. We propose a novel class-imbalanced SSL algorithm called class-distribution-mismatch-aware debiasing (CDMAD). For each iteration of training, CDMAD first assesses the classifier's biased degree towards each class by calculating the logits on an image without any patterns (e.g., solid color image), which can be considered irrelevant to the training set. CDMAD then refines biased pseudo-labels of the base SSL algorithm by ensuring the classifier's neutrality. CDMAD uses these refined pseudo-labels during the training of the base SSL algorithm to improve the quality of the representations. In the test phase, CDMAD similarly refines biased class predictions on test samples. CDMAD can be seen as an extension of post-hoc logit adjustment to address a challenge of incorporating the unknown class distribution of the unlabeled set for re-balancing the biased classifier under class distribution mismatch. CDMAD ensures Fisher consistency for the balanced error. Extensive experiments verify the effectiveness of CDMAD.

Test-time adaptation (TTA) aims to improve model generalizability when test data diverges from training distribution, offering the distinct advantage of not requiring access to training data and processes, especially valuable in the context of large pre-trained models. However, current TTA methods fail to address the fundamental issue: covariate shift, i.e., the decreased generalizability can be attributed to the model’s reliance on the marginal distribution of the training data, which may impair model calibration and introduce confirmation bias. To address this, we propose a novel energy-based perspective, enhancing the model’s perception of target data distributions without requiring access to training data or processes. Building on this perspective, we introduce $\textbf{T}$est-time $\textbf{E}$nergy $\textbf{A}$daptation ($\textbf{TEA}$), which transforms the trained classifier into an energy-based model and aligns the model's distribution with the test data's, enhancing its ability to perceive test distributions and thus improving overall generalizability. Extensive experiments across multiple tasks, benchmarks and architectures demonstrate TEA's superior generalization performance against state-of-the-art methods. Further in-depth analyses reveal that TEA can equip the model with a comprehensive perception of test distribution, ultimately paving the way toward improved generalization and calibration.

Domain adaptation is a critical task in machine learning that aims to improve model performance on a target domain by leveraging knowledge from a related source domain. In this work, we introduce Universal Semi-Supervised Domain Adaptation (UniSSDA), a practical yet challenging setting where the target domain is partially labeled, and the source and target label space may not strictly match. UniSSDA is at the intersection of Universal Domain Adaptation (UniDA) and Semi-Supervised Domain Adaptation (SSDA): the UniDA setting does not allow for fine-grained categorization of target private classes not represented in the source domain, while SSDA focuses on the restricted closed-set setting where source and target label spaces match exactly. Existing UniDA and SSDA methods are susceptible to common-class bias in UniSSDA settings, where models overfit to data distributions of classes common to both domains at the expense of private classes. We propose a new prior-guided pseudo-label refinement strategy to reduce the reinforcement of common-class bias due to pseudo-labeling, a common label propagation strategy in domain adaptation. We demonstrate the effectiveness of the proposed strategy on benchmark datasets Office-Home, DomainNet, and VisDA. The proposed strategy attains the best performance across UniSSDA adaptation settings and establishes a new baseline for UniSSDA.

Vision-Language Models (VLMs) such as CLIP are trained on large amounts of image-text pairs, resulting in remarkable generalization across several data distributions. However, in several cases, their expensive training and data collection/curation costs do not justify the end application. This motivates a vendor-client paradigm, where a vendor trains a large-scale VLM and grants only input-output access to clients on a pay-per-query basis in a black-box setting. The client aims to minimize inference cost by distilling the VLM to a student model using the limited available task-specific data, and further deploying this student model in the downstream application. While naive distillation largely improves the In-Domain (ID) accuracy of the student, it fails to transfer the superior out-of-distribution (OOD) generalization of the VLM teacher using the limited available labeled images. To mitigate this, we propose Vision-Language to Vision - Align, Distill, Predict (VL2V-ADiP), which first aligns the vision and language modalities of the teacher model with the vision modality of a pre-trained student model, and further distills the aligned VLM representations to the student. This maximally retains the pre-trained features of the student, while also incorporating the rich representations of the VLM image encoder and the superior generalization of the text embeddings. The proposed approach achieves state-of-the-art results on the standard Domain Generalization benchmarks in a black-box teacher setting as well as a white-box setting where the weights of the VLM are accessible.

Online continual learning suffers from an underfitted solution due to insufficient training for prompt model update (e.g., single-epoch training). To address the challenge, we propose an efficient online continual learning method using the neural collapse phenomenon. In particular, we induce neural collapse to form a simplex equiangular tight frame (ETF) structure in the representation space so that the continuously learned model with a single epoch can better fit to the streamed data by proposing preparatory data training and residual correction in the representation space. With an extensive set of empirical validations using CIFAR-10/100, TinyImageNet, ImageNet-200, and ImageNet-1K, we show that our proposed method outperforms state-of-the-art methods by a noticeable margin in various online continual learning scenarios such as disjoint and Gaussian scheduled continuous (i.e., boundary-free) data setups.

Unsupervised domain adaptation (UDA) for semantic segmentation aims to transfer the pixel-wise knowledge from the labeled source domain to the unlabeled target domain. However, current UDA methods typically assume a shared label space between source and target, limiting their applicability in real-world scenarios where novel categories may emerge in the target domain. In this paper, we introduce Open-Set Domain Adaptation for Semantic Segmentation (OSDA-SS) for the first time, where the target domain includes unknown classes. We identify two major problems in the OSDA-SS scenario as follows: 1) the existing UDA methods struggle to predict the exact boundary of the unknown classes, and 2) they fail to accurately predict the shape of the unknown classes. To address these issues, we propose Boundary and Unknown Shape-Aware open-set domain adaptation, coined BUS. Our BUS can accurately discern the boundaries between known and unknown classes in a contrastive manner using a novel dilation-erosion-based contrastive loss. In addition, we propose OpenReMix, a new domain mixing augmentation method that guides our model to effectively learn domain and size-invariant features for improving the shape detection of the known and unknown classes. Through extensive experiments, we demonstrate that our proposed BUS effectively detects unknown classes in the challenging OSDA-SS scenario compared to the previous methods by a large margin. The code is available at https://github.com/KHU-AGI/BUS.

Exemplar-free Class Incremental Learning (EFCIL) aims to sequentially learn tasks with access only to data from the current one. EFCIL is of interest because it mitigates concerns about privacy and long-term storage of data, while at the same time alleviating the problem of catastrophic forgetting in incremental learning.In this work, we introduce task-adaptive saliency for EFCIL and propose a new framework, which we call Task-Adaptive Saliency Supervision (TASS), for mitigating the negative effects of saliency drift between different tasks. We first apply boundary-guided saliency to maintain task adaptivity and plasticity on model attention. Besides, we introduce task-agnostic low-level signals as auxiliary supervision to increase the stability of model attention. Finally, we introduce a module for injecting and recovering saliency noise to increase the robustness of saliency preservation. Our experiments demonstrate that our method can better preserve saliency maps across tasks and achieve state-of-the-art results on the CIFAR-100, Tiny-ImageNet, and ImageNet-Subset EFCIL benchmarks.

Zero-shot learning (ZSL) recognizes the unseen classes by conducting visual-semantic interactions to transfer semantic knowledge from seen classes to unseen ones, supported by semantic information (e.g., attributes). However, existing ZSL methods simply extract visual features using a pre-trained network backbone (i.e., CNN or ViT), which fail to learn matched visual-semantic correspondences for representing semantic-related visual features as lacking of the guidance of semantic information, resulting in undesirable visual-semantic interactions. To tackle this issue, we propose a progressive semantic-guided vision transformer for zero-shot learning (dubbed ZSLViT). ZSLViT mainly considers two properties in the whole network: i) discover the semantic-related visual representations explicitly, and ii) discard the semantic-unrelated visual information. Specifically, we first introduce semantic-embedded token learning to improve the visual-semantic correspondences via semantic enhancement and discover the semantic-related visual tokens explicitly with semantic-guided token attention. Then, we fuse low semantic-visual correspondence visual tokens to discard the semantic-unrelated visual information for visual enhancement. These two operations are integrated into various encoders to progressively learn semantic-related visual representations for accurate visual-semantic interactions in ZSL. The extensive experiments show that our ZSLViT achieves significant performance gains on three popular benchmark datasets, i.e., CUB, SUN, and AWA2.

Test-time adaptation (TTA) aims at adapting a model pre-trained on the labeled source domain to the unlabeled target domain. Existing methods usually focus on improving TTA performance under covariate shifts, while neglecting semantic shifts. In this paper, we delve into a realistic open-set TTA setting where the target domain may contain samples from unknown classes. Many state-of-the-art closed-set TTA methods perform poorly when applied to open-set scenarios, which can be attributed to the inaccurate estimation of data distribution and model confidence. To address these issues, we propose a simple but effective framework called unified entropy optimization (UniEnt), which is capable of simultaneously adapting to covariate-shifted in-distribution (csID) data and detecting covariate-shifted out-of-distribution (csOOD) data. Specifically, UniEnt first mines pseudo-csID and pseudo-csOOD samples from test data, followed by entropy minimization on the pseudo-csID data and entropy maximization on the pseudo-csOOD data. Furthermore, we introduce UniEnt+ to alleviate the noise caused by hard data partition leveraging sample-level confidence. Extensive experiments on CIFAR benchmarks and Tiny-ImageNet-C show the superiority of our framework. The code is available at https://github.com/gaozhengqing/UniEnt.

Standard federated learning approaches suffer when client data distributions have sufficient heterogeneity. Recent methods addressed the client data heterogeneity issue via personalized federated learning (PFL) - a class of FL algorithms aiming to personalize learned global knowledge to better suit the clients' local data distributions. Existing PFL methods usually decouple global updates in deep neural networks by performing personalization on particular layers (i.e. classifier heads) and global aggregation for the rest of the network. However, preselecting network layers for personalization may result in suboptimal storage of global knowledge. In this work, we propose FedSelect, a novel PFL algorithm inspired by the iterative subnetwork discovery procedure used for the Lottery Ticket Hypothesis. FedSelect incrementally expands subnetworks to personalize client parameters, concurrently conducting global aggregations on the remaining parameters. This approach enables the personalization of both client parameters and subnetwork structure during the training process. Finally, we show that FedSelect outperforms recent state-of-the-art PFL algorithms under challenging client data heterogeneity settings and demonstrates robustness to various real-world distributional shifts.

Traditional online class incremental learning assumes class sets in different tasks are disjoint. However, recent works have shifted towards a more realistic scenario where tasks have shared classes, creating blurred task boundaries. Under this setting, although existing approaches could be directly applied, challenges like data imbalance and varying class-wise data volumes complicate the critical coreset selection used for replay. To tackle these challenges, we introduce DECO (Dual-Enhanced Coreset Selection with Class-wise Collaboration), an approach that starts by establishing a class-wise balanced memory to address data imbalances, followed by a tailored class-wise gradient-based similarity scoring system for refined coreset selection strategies with reasonable score guidance to all classes. DECO is distinguished by two main strategies: (1) Collaborative Diverse Score Guidance that mitigates biased knowledge in less-exposed classes through guidance from well-established classes, simultaneously consolidating the knowledge in the established classes to enhance overall stability. (2) Adaptive Similarity Score Constraint that relaxes constraints between class types, boosting learning plasticity for less-exposed classes and assisting well-established classes in defining clearer boundaries, thereby improving overall plasticity. Overall, DECO helps effectively identify critical coreset samples, improving learning stability and plasticity across all classes. Extensive experiments are conducted on four benchmark datasets to demonstrate the effectiveness and superiority of DECO over other competitors under this online blurry class incremental learning setting.

Recent compositional zero-shot learning (CZSL) methods adapt pre-trained vision-language models (VLMs) by constructing trainable prompts only for composed state-object pairs. Relying on learning the joint representation of seen compositions, these methods ignore the explicit modeling of the state and object, thus limiting the exploitation of pre-trained knowledge and generalization to unseen compositions. With a particular focus on the universality of the solution, in this work, we propose a novel paradigm for CZSL models that establishes three identification branches (i.e., Multi-Path) to jointly model the state, object, and composition. The presented Troika is an outstanding implementation that aligns the branch-specific prompt representations with decomposed visual features. To calibrate the bias between semantically similar multi-modal representations, we further devise a Cross-Modal Traction module into Troika that shifts the prompt representation towards the current visual content. We conduct extensive experiments on three popular benchmarks, where our method significantly outperforms existing methods in both closed-world and open-world settings.

Open-Set Source-Free Domain Adaptation aims to transfer knowledge in realistic scenarios where the target domain has additional unknown classes compared to the limited-access source domain. Due to the absence of information on unknown classes, existing methods mainly transfer knowledge of known classes while roughly grouping unknown classes as one, attenuating the knowledge transfer and generalization. In contrast, this paper advocates that exploring unknown classes can better identify known ones, and proposes a domain adaptation model to transfer knowledge on known and unknown classes jointly. Specifically, given a source pre-trained model, we first introduce an unknown diffuser that can determine whether classes in space need to be split and merged through similarity measures, to estimate and generate a wider class space distribution, including known and unknown classes. Based on such a wider space distribution, we enhance the reliability of known class knowledge in the source pre-trained model through contrastive constraint. Finally, various supervision information, including reliable known class knowledge and clustered pseudo-labels, optimize the model for impressive knowledge transfer and generalization. Extensive experiments show that our network can achieve superior exploration and knowledge generalization on unknown classes, while with excellent known class transfer. The code is available at https://github.com/xdwfl/UPUK.

Non-exemplar class incremental learning (NECIL) aims to continuously assimilate new knowledge without forgetting previously acquired ones when historical data are unavailable.One of the generative NECIL methods is to invert the images of old classes for joint training. However, these synthetic images suffer significant domain shifts compared with real data, hampering the recognition of old classes.In this paper, we present a novel method termed Dual-Consistency Model Inversion (DCMI) to generate better synthetic samples of old classes through two pivotal consistency alignments: (1) the semantic consistency between the synthetic images and the corresponding prototypes, and (2) domain consistency between synthetic and real images of new classes.Additionally, we introduce Prototypical Routing (PR) to provide task-prior information and generate unbiased and accurate predictions.Our comprehensive experiments across diverse datasets consistently showcase the superiority of our method over previous state-of-the-art approaches. The code will be released.

Few-shot semantic segmentation (FSS) has achieved great success on segmenting objects of novel classes, supported by only a few annotated samples. However, existing FSS methods often underperform in the presence of domain shifts, especially when encountering new domain styles that are unseen during training. It is suboptimal to directly adapt or generalize the entire model to new domains in the few-shot scenario. Instead, our key idea is to adapt a small adapter for rectifying diverse target domain styles to the source domain. Consequently, the rectified target domain features can fittingly benefit from the well-optimized source domain segmentation model, which is intently trained on sufficient source domain data. Training domain-rectifying adapter requires sufficiently diverse target domains. We thus propose a novel local-global style perturbation method to simulate diverse potential target domains by perturbating the feature channel statistics of the individual images and collective statistics of the entire source domain, respectively. Additionally, we propose a cyclic domain alignment module to facilitate the adapter effectively rectifying domains using a reverse domain rectification supervision. The adapter is trained to rectify the image features from diverse synthesized target domains to align with the source domain. During testing on target domains, we start by rectifying the image features and then conduct few-shot segmentation on the domain-rectified features. Extensive experiments demonstrate the effectiveness of our method, achieving promising results on cross-domain few-shot semantic segmentation tasks. Our code is available at https://github.com/Matt-Su/DR-Adapter.

Foundation models encompass an extensive knowledge base and offer remarkable transferability. However, this knowledge becomes outdated or insufficient over time. The challenge lies in continuously updating foundation models to accommodate novel information while retaining their original capabilities. Leveraging the fact that foundation models have initial knowledge on various tasks and domains, we propose a novel approach that, instead of updating all parameters equally, localizes the updates to a sparse set of parameters relevant to the task being learned. We strike a balance between efficiency and new task performance, while maintaining the transferability and generalizability of foundation models. We extensively evaluate our method on foundational vision-language models with a diverse spectrum of continual learning tasks. Our method achieves improvements on the accuracy of the newly learned tasks up to 7% while preserving the pretraining knowledge with a negligible decrease of 0.9% on a representative control set accuracy.

Continual learning has garnered substantial attention within the deep learning community, offering promising solutions to the challenging problem of sequential learning. However, a largely unexplored aspect of this paradigm is its vulnerability to adversarial attacks, particularly those designed to induce forgetting. In this paper, we introduce "BrainWash," a novel poisoning attack specifically tailored to impose forgetting on a continual learner. By adding the BrainWash noise to various baselines, we demonstrate that a trained continual learner can be induced to forget its previously learned tasks catastrophically. A key feature of our approach is that the attacker does not require access to the data from previous tasks and only needs the model's current parameters and the data for the next task that the continual learner will undertake. Our extensive experiments underscore the efficacy of BrainWash, showcasing a degradation in performance across various regularization-based continual learning methods.

Continual learning (CL) aims to empower models to learn new tasks without forgetting previously acquired knowledge. Most prior works concentrate on the techniques of architectures, replay data, regularization, \etc. However, the category name of each class is largely neglected. Existing methods commonly utilize the one-hot labels and randomly initialize the classifier head. We argue that the scarce semantic information conveyed by the one-hot labels hampers the effective knowledge transfer across tasks. In this paper, we revisit the role of the classifier head within the CL paradigm and replace the classifier with semantic knowledge from pretrained language models (PLMs). Specifically, we use PLMs to generate semantic targets for each class, which are frozen and serve as supervision signals during training. Such targets fully consider the semantic correlation between all classes across tasks. Empirical studies show that our approach mitigates forgetting by alleviating representation drifting and facilitating knowledge transfer across tasks. The proposed method is simple to implement and can seamlessly be plugged into existing methods with negligible adjustments. Extensive experiments based on eleven mainstream baselines demonstrate the effectiveness and generalizability of our approach to various protocols. For example, under the class-incremental learning setting on ImageNet-100, our method significantly improves the top-1 accuracy by 3.2\% to 6.1\% while reducing the forgetting rate by 2.6\% to 13.1\%.