Efficient Entropy Coding of Latent Space Coefficients for Point Cloud Geometry Compression using CABAC

Dominik Mehlem, Christian Rohlfing, Jens-Rainer Ohm

Abstract

Over the last years, deep learning-based point cloud geometry compression has emerged as a competitive alternative to conventional octree-based and video-based approaches such as Geometry-based Point Cloud Compression (G-PCC) and Video-based Point Cloud Compression (V-PCC), standardized by MPEG. In particular, methods utilizing variational autoencoders (VAE) with scale hyperprior (HP) have demonstrated strong rate-distortion performance through end-to-end optimized latent entropy modeling. The entropy coding stage relies on learned probabilistic models and arithmetic coding.

In this paper, we propose an entropy coding framework for point cloud geometry latent space coefficients (LSCs) based on Context-based Adaptive Binary Arithmetic Coding (CABAC). The hyperprior-based entropy model is removed from the transmitted bitstream, and the quantized latent coefficients are directly encoded using a newly designed context model with a Star-Template structure. Furthermore, an extensive statistical analysis of the latent space is conducted and shows significant sparsity at low rate points. This property is exploited by introducing a channel-wise masking operation that transmits only significant channels along with a compact mask representation.

In addition, we reassess the rate point configuration of the baseline model and introduce additional low rate operating points, while discarding the highest rate configuration as it is outperformed by lossless G-PCC. Experimental results demonstrate improved BD-rate performance compared to the baseline entropy model when evaluated over relevant rate ranges, with identical reconstruction distortion.

The framework links learned point cloud compression and standardized entropy coding methods, improving coding efficiency while maintaining architectural compatibility with existing VAE-based geometry compression systems. The proposed method achieves significant compression gains, with an average BD-rate reduction of -8.23% in the low rate range and -1.68% overall compared to the baseline.