Early coding unit termination algorithm for video-based dynamic point cloud compression

Abstract

The widespread application of dynamic point clouds is currently hindered by the immense computational complexity of the Video-based Point Cloud Compression (V-PCC) standard. This overhead primarily stems from segmenting massive 3D point clouds into patches, projecting these patches onto 2D geometry and texture maps, and applying the exhaustive quadtree partitioning of High Efficiency Video Coding (HEVC). To mitigate this bottleneck, an adaptive coding unit early termination framework is proposed, which optimizes texture and geometry videos in an independent manner. For texture encoding, an Attention-Integrated Convolutional Neural Network (AM-CNN) is developed to predict Coding Unit (CU) partitioning structures, effectively bypassing redundant rate-distortion calculations. Concurrently, a roughness-aware depth partitioning strategy is introduced for geometry encoding. By mapping 3D surface curvature into 2D spaces, this strategy strategically accelerates mode decisions in unoccupied backgrounds and overly rough regions. Experimental results demonstrate that when integrated into the V-PCC reference model, the proposed framework reduces the overall intra-frame encoding time by an average of 85.99% compared to the standard anchor while maintaining similar perceptual quality, indicating its potential for encoder-side complexity reduction in V-PCC applications.