Feedback Recurrent Autoencoder for Video Compression

Recent advances in deep generative modeling have enabled efficient of high dimensional data distributions and opened up a new horizon for solving compression problems. Specifically, autoencoder based learned image or video solutions are emerging as strong competitors to traditional approaches. In this work, We propose network architecture, on common well studied components, operating low latency mode. Our method yields competitive MS-SSIM/rate performance the high-resolution UVG dataset, among both approaches classical methods (H.265 H.264) rate range interest streaming applications. Additionally, we provide an analysis existing through lens their underlying probabilistic graphical models. Finally, point out issues with temporal consistency color shift observed empirical evaluation, suggest directions forward alleviate those.