Hardware-aware motion estimation via low-resolution motion hints

Abstract

This paper presents an algorithm that achieves high quality video compression with low memory bandwidth of reference frame data and latency due to computation in motion estimation for screen content. Efficiency is attained by content-adaptive placement of the search windows within the reference frames. In our scheme, the center location of the search window is decided by k most prominent motion vectors under a low resolution pre-analysis of the video content. The algorithm leverages the motion hints obtained during pre-analysis to improve encoding efficiency, while keeping implementation complexity and power budget in an acceptable range. Experimental results show that without increasing the size of the search window when large motion is present, it is still possible to capture the motion and achieve within 1.3 dB BDPSNR compared to the HEVC Test Model HM through smart placement of the search window.