PV-RCNN++: semantical point-voxel feature interaction for 3D object detection

Large imbalance often exists between the foreground points (i.e., objects) and background in outdoor LiDAR point clouds. It hinders cutting-edge detectors from focusing on informative areas to produce accurate 3D object detection results. This paper proposes a novel network by semantical point-voxel feature interaction, dubbed PV-RCNN++. Unlike most of existing methods, PV-RCNN++ explores semantic information enhance quality detection. First, segmentation module is proposed retain more discriminative keypoints. Such will guide our integrate object-related point-wise voxel-wise features pivotal areas. Then, make voxels interact efficiently, we utilize voxel query based Manhattan distance quickly sample around reduce time complexity O(N) O(K), compared ball query. Further, avoid being stuck learning only local features, an attention-based residual PointNet designed expand receptive field adaptively aggregate neighboring into Extensive experiments KITTI dataset show that achieves 81.60\(\%\), 40.18\(\%\), 68.21\(\%\) mAP Car, Pedestrian, Cyclist, achieving comparable or even better performance state-of-the-arts.