Fast CU Decision Algorithm Based on LightGBM for Geometry Video in V-PCC

In video-based point cloud compression (V-PCC), geometry video with dual frame rates and high resolution has a unique structure of intra and inter-frame cross-coding in the All-Intra (AI) configuration. It is noted the encoding of geometry video is extremely time-consuming. Despite existing research, the AI configuration still lacks applicable fast coding algorithms for geometry video. This paper proposes a fast CU decision algorithm based on the Bayesian optimization of LightGBM for geometry video. Firstly, by analyzing the differences in CU split about the different layers of the geometry video in the AI configuration, we propose a fast CU split scheme that considers the near and far layers separately. Secondly, according to the property of poor spatial continuity of padded geometry video, this paper explores the relationship between multi-directional edge distributions, block properties of segmental smoothing in CUs, and CU split. Meanwhile, for the unique coding architecture of AI configuration, this paper analyzes the effect of interlayer difference images on CU split in far-layer geometry video. Finally, to reduce the amount of sample data, this paper selects LightGBM to generate the training model and applies Bayesian optimization to determine its optimal hyperparameters. The experiment results indicate the benefits of the proposed method, where it saves encoding time by 40.13%, and the Geom.BD TotalRate for D1 and D2 loss is only 0.03% and 0.08%, respectively.