PV-RCNN++: Point-Voxel Feature Set Abstraction With Local Vector Representation for 3D Object Detection

3D object detection is receiving increasing attention from both industry and academia thanks to its wide applications in various fields. In this paper, we propose the Point-Voxel Region-based Convolution Neural Networks (PV-RCNNs) for 3D object detection from point clouds. First, we propose a novel 3D detector, PV-RCNN, which consists of two steps: the voxel-to-keypoint scene encoding and keypoint-to-grid RoI feature abstraction. These two steps deeply integrate the 3D voxel CNN with the PointNet-based set abstraction for extracting discriminative features. Second, we propose an advanced framework, PV-RCNN++, for more efficient and accurate 3D object detection. It consists of two major improvements: the sectorized proposal-centric strategy for efficiently producing more representative keypoints, and the VectorPool aggregation for better aggregating local point features with much less resource consumption. With these two strategies, our PV-RCNN++ is more than 2x faster than PV-RCNN, while also achieving better performance on the large-scale Waymo Open Dataset with 150m * 150m detection range. Also, our proposed PV-RCNNs achieve state-of-the-art 3D detection performance on both the Waymo Open Dataset and the highly-competitive KITTI benchmark. The source code is available at https://github.com/open-mmlab/OpenPCDet.