The paper presents DistillBEV, a strategy to boost multi-camera bird’s-eye-view (BEV) based 3D object detection in the context of autonomous driving. The authors observe a performance gap between multi-camera BEV and LiDAR-based 3D object detection, largely due to LiDAR’s ability to capture accurate depth and geometry measurements. To bridge this gap, the authors propose training a BEV-based detector to imitate the features of a well-trained LiDAR-based detector. They also propose a balancing strategy to ensure the BEV-based detector focuses on learning crucial features from the LiDAR-based detector. Their experiments show that DistillBEV significantly improves the performance of multi-camera BEV-based detectors.
Publication date: 26 Sep 2023
Project Page: https://arxiv.org/abs/2309.15109
Paper: https://arxiv.org/pdf/2309.15109
