A Fast Adaptive Motion Vector Resolution Algorithm for AVS3

Abstract

Audio Video Coding (AVS3) is one of the latest generation of video coding standards and has significantly improved the coding performance compared to its predecessor. Its efficiency is achieved by the more flexible block partitioning and new encoding tools. Among them, Adaptive Motion Vector Resolution (AMVR) is one of the tools providing efficient coding for motion vectors. It allows motion vector differences to be coded in five resolutions to reduce the encoding bits. However, it’s costly to select the optimal resolution by exhaustively performing motion estimation and rate-distortion optimization. Therefore, we propose our fast AMVR algorithm to select the most probable resolution by utilizing spatial and temporal information. Experimental results demonstrate our algorithm reduces the overall AMVR encoding time by 41%, 32%, and 53% with coding loss of 0.25%, 0.19%, and 0.27% under the low-delay B, low-delay P, and random access configuration.