Distributed Markov decision process in cooperative peer recovery for WWAN multiview video multicast

Abstract

Error resilient video multicast over Wireless Wide Area Networks (WWAN) is difficult because of unavoidable packet losses and impracticality of retransmission on a per packet, per client basis due to the well-known NAK implosion problem. In response, Cooperative Peer-to-peer Repair (CPR) calls for multi-homed devices listening to the same video multicast to locally exchange received WWAN packets via a secondary network like ad hoc Wireless Local Area Network (WLAN) to alleviate individual WWAN packet losses. When videos of the interested 3D scene are captured by multiple closely spaced cameras, each video can be encoded into a separate video stream and transmitted on its own WWAN multicast channel. Clients can then switch observation viewpoints periodically by simply re-subscribing to different WWAN multicast channels- a scenario called interactive multiview video streaming (IMVS). IMVS complicates the CPR WWAN loss recovery process, however, since neighbors of a loss-stricken peer can now be watching different views. In this paper, we optimize the decision process for individual peers during CPR for recovery of multiview video content in IMVS. In particular, for each available transmission opportunity, a peer decides-using Markov decision process as a mathematical formalism-whether to transmit, and if so, how the CPR packet should be encoded using structured network coding (SNC). A loss-stricken peer can then either recover using received CPR packets of the same view, or using packets of two adjacent views and subsequent view interpolation via image-based rendering. Experiments show that decisions made using our proposed MDP outperforms decisions made by a random scheme by at least 1.8dB in PSNR in received video quality in typical network scenario.