NV2P-RCNN: Feature Aggregation Based on Voxel Neighborhood for 3D Object Detection

In this paper, we propose a two-stage framework based on voxel neighborhood feature aggregation for 3D object detection in autonomous driving, named Neighbor Voxels to Point-RCNN (NV2P-RCNN). The point representation of point clouds can encode refined features, and the voxel representation provides an efficient processing framework, so we take advantage of both point representation and voxel representation of the point cloud in this paper. In the first stage, we add point density to the voxel feature encoding and extract voxel features by a 3D sparse convolutional network. In the second stage, the features of the raw point cloud are extracted and fused with the voxel features. To achieve the fast aggregation of voxel-to-point features, we design a neighbor voxels query method named NV-Query to find neighbor voxels directly through the voxel spatial coordinates of the points. The results on the KITTI and ONCE datasets show that NV2P-RCNN achieves higher detection precision compared with other existing methods.