New results in end-to-end image and video compression by deep learning

Abstract

Expanding ubiquity of high-resolution digital video over the Internet calls for better compression methods to enable streaming with higher compression efficiency and lower latency. Recently, important gains have been achieved in learned image compression by using end-to-end learned models. However, these improvements haven't been fully leveraged in video compression. This paper aims to improve upon work proposed by Lu et al. in CVPR 2019, which has been claimed to outperform conventional video codecs in terms of PSNR and provide some implementation details that are absent in the original paper. Ultimately, we show that modeling latent symbols by Laplacian distribution outperforms the Gaussian assumption used in the original work and also demonstrate in a repeatable fashion that our learned model is superior to x264 video codec in terms of PSNR over a range of compression rates measured by bit-per-pixel.