Reinforcing LiDAR-Based 3D Object Detection with RGB and 3D Information

LiDAR-based 3D object detection is efficient for autonomous driving because high accuracy LiDAR information is extremely useful for 3D proposals generation and 3D boxes regression. However, some background and foreground objects may have similar appearances in point clouds. Therefore the accuracy of LiDAR-based 3D object detection is hard to be improved. In this paper, we propose a three-stage 3D object detection method called RGB3D to reinforce LiDAR-based 3D object detection by using an RGB-D classifier with a 3D classifier in parallel. We also apply proper training method to improve the performance of the added classifiers. The 3D classifier is trained by using higher IoU threshold with refined 3D information, and the RGB-D classifier is trained with resized 2D RoIs projected from refined 3D boxes. Extensive experiments are conducted on the KITTI object detection benchmark. The results show that the proposed method is effective.