Question of the architecture (paper)

[handale88]

I wonder about the CPM architecture. I downloaded your code and check out that this works so well. However I wonder why you used Pose Machine architecture for this not inserting more layers for one stage. Also about the intermediate supervision, I guess convolution neural network is good because of not checking the intermediate loss calculation and just checking the loss at last. I read your paper however I still can't fully understand about this. If you want to solve the vanishing gradient problem, you could use ReLU or other solution, i guess..? Please explain about the architecture if I miss something very important in here.
According to the paper, one of the contributions of this paper is that directly operating on belief maps without the need for explicit graphical model-style inference. If I understand it right, please let me know why this is important.
If this is not the adequate place for the paper question, please let me know where to ask question about the paper.

thank you in advance!