Joint Learned and Traditional Video Compression for P Frame

Abstract

In this paper, we propose a joint learned and traditional video compression framework for the P frame track on learned image compression hosted at CVPR2020. The main difference between video compression and image compression is that the former has high degree of similarity between the successive frames which can be utilized to reduce the temporal redundancy. Therefore, we first introduce a decoder-side template-based inter prediction method as an efficient way to obtain reference blocks without the need to signal the motion vectors. Secondly, a CNN post filter is proposed to suppress visual artifacts and improve the decoded image quality. Specifically, the spatial and temporal information is jointly exploited by taking both the current block and similar block in reference frame into consideration. Furthermore, an advanced SSIM based rate-distortion optimization model is proposed to achieve best balance between the coding bits and the decoded image quality. Experimental results show that the proposed P frame compression scheme achieves higher reconstruction quality in terms of both PSNR and MS-SSIM.