Improving Weather-based OOD Generalisation in Lidar-based Object Detection Models via Adversarial Training

Deep learning models are shown to be effective in point-cloud based object detection tasks, such as interpreting data from lidar sensors collected in on-the-road environments. Key to the use of deep learning in safety-critical applications, such as perception in driverless cars, are assurances on accuracy in a wide range of conditions. Previously, point cloud-based object detection models have performed poorly in on-the-road environments when subject to adverse weather, such as rain, snow, and fog. This stems largely from a lack of broad ranging data during training; when collected naturally, adverse weather data is under-represented in training sets. Previous works have tackled this problem using costly simulations to augment training sets with synthetic adverse weather data. In this paper, we propose to improve adverse weather performance in on-the-road, lidar-based object detection tasks through simulation-free adversarial training. We perform benchmarks using PointRCNN against a state-of-the-art simulation-based approach and improve adverse weather performance by up to 6%.