Review for NeurIPS paper: Unsupervised Learning of Object Landmarks via Self-Training Correspondence