Self-supervised Video Representation Learning by Uncovering Spatio-temporal Statistics.