Monocular 3D Object Detection Based on Pseudo-LiDAR Point Cloud for Autonomous Vehicles

Pseudo-LiDAR point clouds are generated from monocular image. Compared with the point cloud from LiDAR, it can provide denser data. Due to the inaccuracy of depth estimation, there is still a performance gap between the pseudo-LiDAR point cloud and the LiDAR one. In this paper, we propose an approach to eliminate the performance degradation caused by deviation of depth estimation and realize 3D object detection based on pseudo-LiDAR point cloud. Comparative results on KITTI 3D benchmark illustrate that in contrast to other methods, our scheme can achieve more reliable performance on both object localization and shape estimation.