Mode Duplication Based Multiview Multiple Description Video Coding

Abstract

Compression ability is most concerned for multiview video (MVV) transmission system because of its massive amount of data. To improve coding efficiency, the Joint Video Team (JVT) standardization body has developed a joint multiview video coding model (JMVC), in which both intra-view and inter-view prediction techniques are exploited to yield a better coding gain. Therefore, how to prevent from error propagation has become a critical issue in multiview video coding (MVC). Error concealment methods for MVC have been widely studied in recent years, but little research has been conducted on error resilience for MVC. Multiple description coding (MDC) provides a promising solution for robust data transmission over error-prone channels, and has found many applications in monoview video communications. However, these MDC frameworks are not applicable to MVC because its prediction structure involves inter-view prediction. To develop an efficient and robust MVC scheme, a novel MDC algorithm for JMVC based on the mode duplication strategy is proposed in this paper. The input MVV sequence is firstly sub sampled in both horizontal and vertical directions, forming four subsequences, X1p, X1d, X2p, and X2d. Then, X1p and X1d are paired to form description 1, and X2p and X2d are grouped to form description 2, respectively. Secondly, X1p and X2p are directly encoded by separate JMVC encoders. Meanwhile, X1d/X2d adopts the best modes and prediction vectors (PVs) of X1p/X2p in corresponding (same spatial) locations to perform prediction coding. Consequently, neither code for best modes and PVs nor time load for mode decision is needed while coding X1d and X2d. Only coding for the prediction errors is required. Because subsequences in the same description are closely resembled each other, the extra prediction errors introduced by this best mode and PV reuse are negligible. Therefore, the bit rate and computational cost for coding X1d and X2d are greatly reduced. The proposed algorithm has been integrated into JMVC 6.0 and experimented on multiple MVV test sequences. The experimental results have shown that the proposed algorithm outperforms the state-of-the-arts of MDC for MVV and stereoscopic video, achieving improvements of 0.5-3dB in central decode and 0.5-3.5dB in side decode at the same bit rate over a wide range from 500kbps to 6000kbps. Comparing with original JMVC, the proposed algorithm saves about 40% encoding time in average.