Stereo Point Cloud Refinement for 3D Object Detection

3D object detection has shown advantages over its 2D image based counterpart. This paper proposed a new pipeline to utilize the left and right consistence check on disparity map for stereo point clouds-based 3D object detection. Unlike existing pipeline directly project the depth map to the 3D space, the proposed pipeline first use the left and right consistence to filter out the bad pixels in the disparity map before the projection to stereo point clouds. Experimental results show that by eliminating those bad points, the proposed pipeline can achieve better performance in 3D object detection tasks. Moreover, due to the reduced number of points, the computation cost of 3D object detection can be significantly reduced.