Online learning for quality-driven unequal protection of scalable video

Abstract

Video packet losses affect perceived video quality non-uniformly due to several factors related to video encoding such as inter-frame coding and motion compensation as well as due to psycho-visual perception of natural scenes with unequal motion. This motivates protecting video packets unequally based on their loss visibility. This paper proposes an adaptive online algorithm for unequal error protection driven by two key motivations: On one hand, for real-time video, where a video sequence is not pre-encoded, an offline approach is infeasible and determining the unequal protection levels to maintain a target video quality level must be performed online. On the other hand, an online approach enables adapting to scene changes as well as changes in video temporal and spatial characteristics. The proposed online algorithm uses local linear regression to learn the mapping between packet losses from each scalable video layer and quality degradation without assuming an underlying statistical model. The notion of locality captures the similarity in video scene characteristics as well as proximity in time. The algorithm provably guarantees an average target video quality level and converges rapidly to a stable solution. Furthermore, it provides a bias/variance tradeoff between factual estimation of loss visibility and fine adaptation to the changing video temporal characteristics.