Reinforcement Learning-Based Rate Adaptation for Point Cloud Streaming

With the ongoing desire to achieve six degrees of freedom in immersive virtual worlds, and the technological advances in point cloud capturing techniques such as terrestrial laser scanning (TLS), point cloud streaming has lately been an emerging and fast growing research topic. Existing Dynamic Adaptive Streaming over HTTP (DASH)-like approaches require the point clouds to be fully available before they can be streamed. In this paper, we consider a delay-sensitive scenario where the point cloud data should be transmitted in real-time from a laser scanner, while the laser scan is still ongoing. To ensure real-time communication and high quality of the transmitted data, the sending data rate must be adapted to the available network capabilities. We explore the applicability of reinforcement learning to this rate adaptation problem. Our evaluation results show that our approach outperforms existing rate adaptation heuristics by up to 41.41%.