Local and Global Viewport History Sampling for Improved User Quality of Experience in Viewport-Aware Tile-Based 360-Degree Video Streaming

Generating accurate predictions of a user's region of interest in 360-degree video streaming is a well-known challenging task due to the erratic nature of user head movements. A misprediction in the user's region of interest can greatly impact the Quality of Experience (QoE) for the user depending on the severity. In this paper, to mitigate the severity of mispredictions, we propose a novel viewport extension system in which clustering approaches on past users' viewport history information can create insightful extensions to the current user's predicted region of interest. In addition, we present two system types for how and when the clustering approaches can be applied between client and server for viewport extensions at prediction. Based on simulations on ten 360-degree video datasets projected at small and large equirectangular tiling, we first demonstrate that the local approach of clustering past users into small groups can further increase user QoE compared to the global approach of using all past users' viewport history information. Secondly, we present analysis in which environments a local or global viewport history sampling approach can yield the best performance in user QoE.