Neural Network Modeling of Probabilities for Coding the Octree Representation of Point Clouds

Abstract

This paper describes a novel lossless point cloud<br/>compression algorithm that uses a neural network for estimating<br/>the coding probabilities for the occupancy status of voxels,<br/>depending on wide three dimensional contexts around the voxel<br/>to be encoded. The point cloud is represented as an octree, with<br/>each resolution layer being sequentially encoded and decoded<br/>using arithmetic coding, starting from the lowest resolution, until<br/>the final resolution is reached. The occupancy probability of<br/>each voxel of the splitting pattern at each node of the octree<br/>is modeled by a neural network, having at its input the already<br/>encoded occupancy status of several octree nodes (belonging to<br/>the past and current resolutions), corresponding to a 3D context<br/>surrounding the node to be encoded. The algorithm has a fast<br/>and a slow version, the fast version selecting differently several<br/>voxels of the context, which allows an increased parallelization by<br/>sending larger batches of templates to be estimated by the neural<br/>network, at both encoder and decoder. The proposed algorithms<br/>yield state-of-the-art results on benchmark datasets. The imple-<br/>mentation will be available at https://github.com/marmus12/nnct