Exploiting the layer correlation to improve DASH scheduling with scalable video coding

Abstract DASH is a promising video streaming solution that gracefully deals with network throughput fluctuations as well as terminal heterogeneity. In DASH, an MPD file describes the available video qualities and their average bitrates. However, the coarse-grained average bitrate is insufficient for an accurate client-side scheduler. The bitrate overestimate or underestimate of bandwidth requirements for segments to be fetched could result in suboptimal scheduling decisions and possibly drain the playback buffer. In this paper, we first made the observation that with SVC, video segments across different layers exhibit high size correlation. We then showed that enhancement layer segment size can be predicted by the corresponding base layer segment with relatively high accuracy. Based on this observation, we enhanced existing rate-based and buffered-based adaptive bitrate (ABR) algorithms with a size predictor, which is orthogonal to existing adaptation logics. Results show that augmented by the size predictor, the average playback bitrate can be improved by up to 20% and playback stalls can also be significantly reduced. Our results demonstrated that similar to the bandwidth predictor, a size predictor should also be added to SVC-based ABR algorithms to increase the performance gain.