Viewport prediction for 360° videos: a clustering approach

Abstract

An important component for viewport-adaptive streaming of 360° videos is viewport prediction. Increasing viewport prediction horizon enables the client to prefetch more chunks into the playback buffer. Having longer buffer results in less rebuffering under fluctuating network conditions. We analyzed the recorded viewport traces of viewers who watched various 360° videos. We propose a clustering-based viewport prediction method that incorporates viewport pattern information from previous video streaming sessions. For several videos, specifically those with well-defined region of interest, the proposed approach increases the viewport prediction horizon and/or prediction accuracy.