Low-complexity Deep Video Compression with A Distributed Coding Architecture

Xinjie Zhang, Jiawei Shao, Jun Zhang

Keywords: Deep video compression, distributed coding, low encoder complexity

TL;DR: We design the first end-to-end distributed deep video compression framework based on the distributed coding paradigm, which outperforms traditional distributed video codecs and achieves competitive performance with H.264.

Abstract: Prevalent video compression methods follow a $predictive\;coding$ architecture that relies on a heavy encoder to exploit the statistical redundancy, which makes it challenging to deploy them on resource-constrained devices. Meanwhile, as early as the 1970s, distributed source coding theory, namely, Slepian-Wolf and Wyner-Ziv theorems, has indicated that efficient compression of correlated sources can be achieved by exploiting the source statistics at the decoder only, with the help of effective side information (SI). This has inspired a $distributed\;coding$ architecture that is promising to reduce the encoder complexity. While there have been some attempts to develop practical distributed video coding systems, traditional methods suffer from a substantial performance gap to the predictive coding architecture. Inspired by the recent successes of deep learning in enhancing image and video compression, we propose the first end-to-end distributed deep video compression (Distributed DVC) framework with neural network-based modules that can be optimized to improve the rate-distortion performance. A key ingredient is an effective SI generation module at the decoder, which helps to effectively exploit the inter-frame correlation without computation-intensive encoder-side motion estimation and compensation. Experiments show that Distributed DVC significantly outperforms conventional distributed video coding methods and H.264. Meanwhile, it enjoys $6\sim7$ times encoding speedup against DVC with only 1.61\% increase in the bitrate for 1080P test videos on the UVG dataset.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2303.11599/code)