CVPCG: Centrosymmetric Virtual Point Cloud Generation For 3D Object Detection

Abstract

Although point-based 3D object detection methods have made great progress, these methods still have the problem of low precision due to the incomplete information collected by lidar. To solve this problem, this paper proposes a virtual point cloud generation architecture for 3D object detection. Firstly, the point cloud is converted into voxel representation and input to the voxel transformer to extract local and global features, and then the 3D features are converted into 2D features for one-stage low confidence threshold prediction. Secondly, the points inside each prediction box will generate virtual points through the centrosymmetric method, and then the prediction results of the high confidence threshold of the second stage are generated through a series of 3D sparse convolution and 2D backbone network. Finally, various features are fed into the RoI-grid Pool to generate confidence results. Experimental results on KITTI and Waymo Open dataset show that our method is effective, and the precision have significant advantages compared to other methods.