Efficient Deep Super-Resolution of Voxelized Point Cloud in Geometry Compression

Point cloud compression is an essential task for practical applications using point clouds. Most of the previous approaches rely on octree compression which involves voxelization in the coding itself. Distortions derived from voxelization can be reduced without increasing the bitrate by postprocessing. In this article, we propose a super-resolution method for a decoded voxelized point cloud as a postprocessing step in the geometry compression. The proposed method increases the resolution of the voxelized point cloud by predicting the occupancy of higher resolution voxels than those used to compress the original point cloud. For efficient prediction, we propose a deep neural network for super-resolution based on sparse convolution. It can be highly efficient even for a large point cloud since the network applies convolution only to nonempty space. The proposed method predicts the occupancies represented by continuous values for each point and estimates the binary occupancies through a thresholding procedure. We design a dynamic threshold to ensure that at least one of all voxels is predicted to be occupied in order to prevent the generation of regions with missing points. We also introduce an occupancy prediction method to address the sparsity of high-resolution occupied voxels. Experiments on the outdoor and indoor datasets demonstrate the effectiveness of the proposed method.