Learning View-Disentangled Human Pose Representation by Contrastive Cross-View Mutual Information Maximization.