Lossless Compression of Point Cloud Sequences Using Sequence Optimized CNN Models

In this paper we propose a new paradigm for encoding the geometry of dense point cloud sequences, where convolutional neural network (CNN), which estimates distributions, is optimized on several frames sequence to be compressed. We adopt lightweight CNN structures, perform training as part process and parameters are transmitted bitstream. The newly proposed scheme operates octree representation each cloud, consecutively resolution level. At every level, voxel grid traversed section-by-section (each section being perpendicular selected coordinate axis), in section, occupancies groups two-by-two voxels encoded at once single arithmetic coding operation. A context conditional distribution defined group based information available about occupancy neighboring current lower layers octree. probability mass functions patterns all from one four phases. phase, contexts updated with previous phase probabilities parallel, providing reasonable trade-off between parallelism processing informativeness contexts. time comparable spent remaining steps, leading competitive overall times. bitrates encoding-decoding times compare favorably those recently published compression schemes.