VPCNet: Voxel-Point Cascade for 3D Object Detection

Most current 3D object detectors employ the sophisticated point-voxel fusion mechanisms to obtain high-quality feature representation. However, they rarely consider the hierarchical relationship between points and voxels. We propose a novel detection network, called VPCNet, which fuses voxels and raw points utilizing a cascaded structure to optimize proposals hierarchically. The proposed Hierarchical Head is a two-level detection head, including voxel level and point level. The first level utilizes voxel representation to predict relatively rough boxes. The second level exploits the Corner Point Transformer to model the spatial context relationship between the original points and the previous level boxes to obtain more accurate detection results. A Quadratic Weighted Decoder of Corner Point Transformer is developed with standard transformer decoder to reduces the adverse effect of sparsity on learned features. This design combines computationally efficient voxel-based methods with point-based methods which possess more complete spatial information. Our method outperforms most state-of-the-art 3D object detectors on both KITTI and Waymo Open Datasets.