We thank all the reviewers for excellent questions and many relevant remarks. Thank you for this remark. One of the reason for this is that our method produces interpretations directly in terms of the input features. Thank you for pointing this out, we agree that faithful is not best. This is not the case for local models such as LIME.

"motivation [...] very convincing and perfectly pitched to the reader" We believe that this will spur follow-up work benefitting both of these promising research directions. We have trained models on the ShapeNet "benches" class--please see qualitative We note that 2D results (Sec. Figure 1 of DeepSDF--see qualitative result in (c)--with no further fine-tuning or heuristics. We will add experiments and comparisons with further classes to the final manuscript. We will discuss DISN in-depth. We benchmark against this architecture (see submission Table 3, Figure 1).

First provide a summary of the paper, and then address the following criteria: Quality, clarity, originality and significance. This paper proposes a novel approach to human pose estimation, consisting of a deep convolutional network for part detection and a higher-level spatial model that is motivated as a graphical model, but actually incorporated into the overall deep network as a particular sub-net that has the plausible interpretation of performing a single round of message passing. The system is trained in three steps. In the first two steps, the deep convolutional part detector and the spatial model are trained individually (the spatial message passing network uses the heat map output of the part detector), while in the third step, the unified network is jointly trained via back propagation. Even though the convolutional part detector alone is already a state-of-the-art system, the spatial model is shown to improve results considerably, with even further improvements gained via the joint training procedure.