Motion Estimation Algorithm Based on Decoded Residual Compensation: A Proof-Of-Concept

Lee Sze Foo, Anissa Mokraoui, Fangchen FENG, Yoong Choon Chang

Abstract: In most video coding standards, the reduction of temporal redundancy in a video is based on the traditional block-matching algorithm (BMA). It first estimates the motion vectors that minimize the distortion between the original image and its predicted version. The difference between these two images, i.e. residual image, is then encoded and its decoded version compensates the predicted image. This paper proposes an algorithm that estimates the motion vectors while taking into account the impact of the decoded residual image on the quality of the compensated image. This algorithm provides a higher PSNR for a given bit rate compared to the traditional method. This proof-of-concept shows the importance of taking into account the compensated image in the motion vector estimation process and should help in the design of solutions based on deep neural networks.