Few-Shot Online Learning for 3D Object Detection in Autonomous Driving

For autonomous driving, the performance of 3D object detection is limited by offline training, and these methods usually lack the adaption ability for long-term autonomy, which leads to significant performance degeneration across different scenarios, i.e. domain shift. This paper proposes a few-shot online learning method to transfer knowledge from 2D images to 3D point clouds. In particular, the point cloud clusters are automatically labeled by the 3D-2D projection and 3D object tracking, and the learning strategy allows the classifier to learn multiple classes with limited samples in a short period of time. The final 3D detection results are obtained from the fusion of the online learning 3D detector and an end-to-end 3D detector. Experimental results on the KITTI dataset demonstrate the effectiveness of our system compared to the baseline methods.