SemiPoint: Generalizing Cross-Scene Point Cloud Video Streaming With Semi-Supervised Learning and Residual-Aware Adaptation

Viewport prediction algorithms are shedding new light on point cloud video (PCV) streaming. Most existing methodologies are trained with labeled frames (supervised learning) to reduce bandwidth consumption. However, the fully supervised paradigm requires labor-intensive video labeling, struggles to generalize to unfamiliar scenes, and thus produces noisy bitrate allocation outputs. In this study, we propose SemiPoint, a cross-scene PCV streaming framework that features a semi-supervised viewport prediction module and a residual-augmented deep reinforcement learning (DRL)-based bitrate adaptation module. The viewport prediction module employs a semi-supervised architecture that enhances scene generalization by exploiting unlabeled frames through unsupervised constraints. Furthermore, the DRL-based bitrate adaptation module incorporates a residual model that dynamically corrects abrupt viewport shifts through real-time residual compensation. Extensive experimental evaluations demonstrate that SemiPoint achieves superior performance compared to fully supervised approaches with limited labeled datasets. It demonstrates enhanced generalization capabilities in changing scenes and delivers more reliable bitrate adaptation in scenarios involving sudden head/body movements.