Distributed Multiple Description Video Coding

Multiple description video coding (MDVC) has long been recognized as a promising technique to enhance the robustness of video streams. A multiple description video encoder generates multiple video streams. Each stream can be decoded independently with acceptable reconstruction quality. Several streams can be jointly decoded to yield a better quality. Currently, almost all the MDVC schemes use multiple prediction loops, which succeed in achieving high coding efficiency on ideal multiple description (MD) channels, where each stream is either received intact, or lost totally. However, these algorithms suffer from severe predictive mismatch problem on random packet loss channels, where the side information at the decoder will most probably be different from that at the encoder. This paper proposes Distributed Multiple Description Video Coding (DMDVC), where distributed source coding is used to achieve compression instead of predictive coding, in order to solve the predictive mismatch problem. Since distributed source coding makes use of the statistical correlation between the source and the side information instead of the sample paths of both, successful decoding may still be achieved even if the side informations of both the encoder and the decoder are different. Thereby, High robustness of compressed video streams is achieved.An information theoretical problem named multiple description coding in the Wyner-Ziv setting can be viewed as a model for coding one frame in DMDVC. Therefore, code constructions for the problem are presented first. Both theoretical analysis and experiment results show that the performance loss of our code constructions is 2~3dB compared to the theoretical bound. Our MD video codec is based on our code construc- tions as well as state-of-art distributed video coding algorithms. Since no prediction loops are used to achieve inter-frame compression, our video schemes do not suffer from the notorious predictive mismatch problem. Experiment results demonstrate significant improvement of our algo- rithms over classical ones based on prediction loops at high packet loss rates.