Residue-free video coding with pixelwise adaptive spatio-temporal prediction

The authors introduce residue-free video coding, in which motion-compensated predictions from surrounding frames and spatial predictions from the current frame are combined adaptively on a pixel-by-pixel basis. The consequence is that residue frames, blocks or regions are never explicitly formed. The authors describe a practical embodiment of a residue-free coder- temporal prediction trees - in which the local adaptation is conditioned frame to frame by a control parameter derived from global motion statistics. Using fixed-block-size motion compensation, the resulting coder is competitive with conventional residue-based compression, and at higher data rates is able to outperform H.264/AVC for high-activity sequences.