ViewCLR: Learning Self-supervised Video Representation for Unseen Viewpoints.