Frame synthesis for video compression

Traditional video compression techniques heavily rely on the concept of motion compensation to predict frames in a video given previous or future frames. With the recent advances in artificial intelligence powered techniques, new approaches to video coding become apparent. We describe a video compression scheme that employs neural network based image compression and frame generation. The proposed method encodes key frames (I frames) as still images and entirely skips compression and signalling of intermediate frames (S frames), which are conversely synthesized on the decoder side exclusively using I frames. Varying complexities of motion can occur within a sequence, which can have a strong impact on the quality of the generated frames. In order to address this challenge, we propose to let the encoder dynamically adjust the group of pictures (GOP) structure. This adjustment is performed based on the quality of predicted S frames. The achieved performance of the proposed method suggests that entirely skipping coding and instead synthesizing frames is promising and should be considered for future developments of learning based video codecs.