Questions about the 2D Pose Sub-Network

[nsturis]

Hi,

Very exciting paper! We're trying to replicate your implementation in PyTorch and are running into difficulties regarding the 2d Pose Sub-Network. We were wondering if you could help with our understanding, if you have the time.

Firstly, we're doing this based on your descriptions in your paper and from the prototxt's in your repo. If the questions below are easily answered by installing Caffe and trying things out there, please let us know, but we still think there are som points of confusion:

There are two different descriptions of the 2D Pose Sub-Network in your paper. In section 3.1:

Which seems to corrospond to the pose_train_test_from_release.prototxt, to the original CPM paper and your pre-trained CPM-caffemodel. However, later on in section 3.4 you write:

Which seems to corrospond to the train_test_cpm.prototxt.

The number of stages in the CPM is also different between the two prototxts, with the first one containing one refinement stage, but the second containing none. In your paper you don't specify any refinement stages for the 2D Pose Sub-Network. Are refinement stages recommended or not? Did you use that in your paper? And, is the 2d joint position regressed with argmax (3.1) or an extra fully connected layer (3.4)?

Other questions if you have time:

1. Why do you specify in section 3.2 that there is a max-pool layer in the 2d-to-3d pose transformer after the two convolutional layers, when there's no layer specified in the prototxt?

2. When looking at deploy.prototxt the sizes of the convolutional layers also seem to differ from the paper. You specify 46 -> 23 -> 11, but in the prototxt it is 46 -> 22 -> 10.

Thank you for your time.

Kind regards,
Noah