convolution feature extractor with a hierarchical, fully differentiable matching layer based on entropy regularized

First of all, we would like to thank the reviewers for their constructive feedback and thorough reviews. In comparison to prior work, our network is more accurate and generalizes better across benchmarks. Table 1 prove that our method is robust to this type of input noise. We will state this point more clearly in the paper and thank the reviewer for the insight. According to [7, Appendix C] this leads to problems for humans in "bent over poses", see also Figure 2 of our Appendix.