3D Object Detection on Voxels in Spherical Coordinate System

LiDAR sensors are commonly used to acquire 3D point cloud data in autonomous driving and robotics. Voxel-based 3D convolutional networks have been used for some time to extract features from point clouds for object detection. However, most of the existing work converts point clouds into voxels in the Cartesian coordinate system. In this paper, we describe how to perform 3D detection on voxels in a spherical coordinate system. We advocate spherical voxelization with the following benefits: (a) Less information loss; (b) Balanced voxelization of points, which also means balanced distribution of computing power to points; (c) No need to drop points randomly, yielding deterministic voxel embeddings. Also, we propose using a logarithmic scale on the range axis to get a more balanced voxelization. Based on such voxelization, we build an anchor-based detection network, SphVoxNet. We show that detection using a spherical coordinate system can reach a comparable or even better performance than using the Cartesian coordinate system.