Splatting-based Motion Context Encoding for Deep Video Compression
Keywords: Video Compression, Motion Compensation, Splatting
TL;DR: In this paper, we propose a methodology to compress motion information in videos more efficiently using forward warping, rather than the conventional backward warping approach.
Abstract: Recent video compression studies aim to compress videos in a more optimal space using deep neural networks. Most of them employ a strategy where they use motion information to warp the previous frame to align with the current frame, and then only compress the information newly appearing in the current frame. While this enhances the compression efficiency of each frame, additional bits are required to compress the motion information alongside it. In this paper, we explore a methodology that improves motion compression by warping previous motions just like frames. However, within the traditional backward warping-based framework, a dilemma arises where the decoded motion is needed to warp the reference motion. To solve this problem, we propose a forward warping-based framework for video compression called SVC (Splatting-based Video Compression). While SVC offers the advantage of enabling the use of motion context, forward warping has several issues compared to backward warping and we propose additional tricks to address these challenges. Intensive experiments on the UVG, HEVC, and MCL-JCV benchmarks demonstrate that motion context encoding through SVC is indeed more effective compared to various methods based on backward warping, including traditional codecs.
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1468
Loading