QoE-driven Mobile 360 Video Streaming: Predictive View Generation and Dynamic Tile Selection

Abstract

In mobile video streaming, 360-degree videos can provide users with immersive and memorable experience. Due to the panoramic and high resolution features, limited bandwidth and stringent latency requirements, the transmission of full high-definition 360-degree video may cause severe stalling, significantly lowering the users' quality of experience (QoE). As the video content seen by the user largely relies on the user's viewing direction and the size of field of view, in this paper, we investigate viewpoint prediction and dynamic tile selection to improve users' QoE for mobile 360-degree video streaming. Specifically, we first design a recurrent neural network integrated with attention mechanism to predict the user's viewpoint in the next video segment. We then propose a dynamic tile-selection method which selects and transmits the tiles that are most likely to be viewed in a segment through online learning. Experimental results based on a real-world dataset show that, the proposed viewpoint prediction neural network and dynamic tile selection method can effectively improve the prediction accuracy and improve the users' QoE.