Online learning for quality-driven unequal protection of scalable video

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