A Study of Motion Coding Schemes for Learned Video Compression

This paper presents a study of motion coding schemes for learned video compression. Most learned video compression systems explicitly signal optical flow maps to characterize motion between video frames for motion compensation. The flow maps, usually of the same size as the video frames, represent a considerable portion of the compressed bitstream. This work studies several schemes to make a non-linear prediction of the flow maps for efficient motion coding. These include signaling an incremental flow map between a coding frame and a motion-compensated frame derived from the flow map predictor. In forming the flow map predictor, we propose a learned motion extrapolation module and a motion forward warping scheme. They are further incorporated into two novel approaches, termed double warping and frame synthesis with motion forward warping, in creating an inter-frame predictor by combining the incremental flow and the flow map predictor. Extensive experiments are conducted to analyze the merits and faults of these variants, and demonstrate their superiority to predictive motion coding and intra motion coding.