Deep Dynamic Point Cloud Attribute Compression Using Dual-Modal Motion Estimation and Spatio-Temporal Conditional Residual Coding

Dynamic point cloud is a sequence of point cloud, providing 3D geometry coordinates, attribute as well as temporal information. However, its massive data volume hinders the data processing, storage and network transmission for 3D applications. To address this problem, we propose an end-to-end Deep Dynamic Point Cloud Attribute Compression (DD-PCAC) to exploit the temporal redundancies in dynamic point cloud. First, we design a unified DD-PCAC framework that performs motion estimation, compensation, and residual encoding in a fully differentiable pipeline. Second, we propose a Dual-Modal Motion Estimation (DMME) module to explore temporal inter-frame redundancies of dynamic point cloud, which includes Dynamic Ball Matching (DBM), Pixel-domain Aggregation (PDA), and Deep Feature Aggregation (DFA) modules. Third, we propose a spatio-temporal conditional residual encoder-decoder to effectively model temporal dependencies of point cloud attribute and improve compression efficiency. Experimental results show that the proposed DD-PCAC achieves an average of 31.57% and 30.33% BD-rate gains compared with the state-of-the-art TSC-PCAC model and the traditional Geometry-based Point Cloud Compression (G-PCC), respectively. Moreover, the encoding and decoding complexities of the proposed DD-PCAC are much lower than the benchmark G-PCC, V-PCC and Sparse-PCAC. The source code and pretrained models are available at: https://github.com/SYSU-Video/DD-PCAC