A Temporal Dependency Model for Rate-Distortion Optimization in Video Coding

Abstract

Video codec heavily relies on motion compensated prediction to achieve compression efficiency. The predictive scheme creates temporal dependency across frames, i.e., the quantization distortion in a current block may propagate through motion compensated prediction and affect the coding efficiency of blocks in subsequent frames. The ability to capture such dependency can potentially improve the rate-distortion optimization for coding performance gains. Prior research work builds block-based motion trajectories and uses the correlations between source pixel blocks in the same motion trajectory to estimate the distortion propagation model. This work premises on the realization that the distortion propagation is also largely related to the quantization effect. A novel temporal dependency model that accounts for both block correlation and the quantization effect is proposed. It is experimentally shown to provide considerable compression gains over the existing competitors.