The authors propose a combination of diffusion distances with deep networks trained for image segmentation. Two of the reviewers felt that there were interesting methodological contributions meriting acceptance, while all three reviewers indicated that baselines in the experiments could be improved.