Implicit Alignment and Long Temporal Context Memory Propagation-Based Video Compression Framework

Abstract

Recently, an efficient video compression framework utilizing contextual compression instead of residual compression has been proposed, but it exhibits certain shortcomings in the generation of contextual information. To address the shortcomings of artifacts appearing when warping reference frames using optical flow in pixel space in the contextual compression framework, as well as the introduction of additional errors by bilinear interpolation, we propose a video compression framework based on implicit alignment and long-term contextual memory propagation. This network performs motion estimation in feature space, utilizing the adaptability of deformable convolution to achieve implicit alignment of video frame features. We also propose a long-term feature propagation network and a long-term context refinement network to accumulate feature information from all previous frames and use memory cells to refine the contextual information through a refinement network. Additionally, our framework removes the autoregressive module, which can only be serialized in the original model.