One key challenge to learning-based video compression is that motion
predictive coding, a very effective tool for video compression, can hardly be
trained into a neural network. In this paper we propose the concept of
PixelMotionCNN (PMCNN) which includes motion extension and hybrid pr