In this document we provide more details about implementation and experiments about different voxelization methods. For z we set 12 bins and range [ 5, 3]. We set 12 bins and range [ 3, 3] for log l, log, w, log h. We use code weight 0.5 for velocity prediction and 1.0 for other bounding box statistics. In the classification head, we set alpha=0.25, gamma=2.0 for Focal Loss [1].

Recent voxel-based 3D object detectors for autonomous vehicles learn point cloud representations either from bird eye view (BEV) or range view (RV, a.k.a.

We thank all the reviewers for their efforts and thoughtful feedbacks. Moreover, we provided in-depth analysis and clear ablation studies to validate our contributions. Our algorithm runs at 8 FPS with a single V100 GPU on Waymo Open Dataset. The experiments are conducted on NuScenes validation set. MB of paramters (18% more) in our proposed CVCNet.