Voxel R-CNN: Towards High Performance Voxel-based 3D Object Detection

Recent advances on 3D object detection heavily rely how the data are represented, i.e., voxel-based or point-based representation. Many existing high performance detectors because this structure can better retain precise point positions. Nevertheless, point-level features lead to computation overheads due unordered storage. In contrast, is suited for feature extraction but often yields lower accuracy input divided into grids. paper, we take a slightly different viewpoint --- find that positioning of raw points not essential and coarse voxel granularity also offer sufficient accuracy. Bearing view in mind, devise simple effective framework, named Voxel R-CNN. By taking full advantage two-stage approach, our method achieves comparable with state-of-the-art models, at fraction cost. R-CNN consists backbone network, 2D bird-eye-view (BEV) Region Proposal Network, detect head. A RoI pooling devised extract directly from further refinement. Extensive experiments conducted widely used KITTI Dataset more recent Waymo Open Dataset. Our results show compared methods, delivers higher while maintaining real-time frame processing rate, speed 25 FPS an NVIDIA RTX 2080 Ti GPU. The code available https://github.com/djiajunustc/Voxel-R-CNN.