Deep Contextual Dynamic Point Cloud Geometry Compression

Existing dynamic point cloud compression methods adopt a predictive coding framework, where the residual between a predicted frame and the current frame is calculated and encoded based on the prediction of the current frame’s point cloud. However, due to the use of residual coding to eliminate inter-frame redundancy, the compression performance in regions with intense motion and rich content details is subpar. To address this limitation, this paper proposes a depth context-based dynamic point cloud geometry compression method using conditional coding instead of traditional predictive coding. Firstly, a feature space multi-scale scene flow network is constructed to calculate the motion vectors of the dynamic point cloud, enhancing the accuracy of motion estimation. Secondly, a context generation module is utilized to construct context information, which is then directly used as a condition for encoding and decoding the current frame’s point cloud. Lastly, by fusing temporal priors with hyperpriors, a context-based entropy model is designed to estimate the conditional probability distribution of the point cloud, thereby improving the efficiency of entropy coding. Evaluated using the BD-rate of rate-distortion curves as the performance metric, the proposed method achieves an average bitrate reduction of 13.62% compared to D-DPCC on the 8iVFB dataset and 15.15% on the MVUB dataset, significantly enhancing compression performance.