How were the ground truth normal maps computed for training the pix2pix network?

[janehwu]

I understand the normal maps are computed in camera space, but could you please elaborate on what the exact transformation is from world space normals to camera space? E.g. in camera space, how are the x,y,z axes defined?

I've been looking at this code, but when I pass a predicted mesh given by the demo code to _render_normal I get a blank image (which suggests these are not the right transformations):

pifuhd/lib/evaluator.py

Line 83 in e47c4d9

def _render_normal(self, mesh, deg):

In the demo code, it looks like the normals are directly predicted by the network, so I'm having trouble deciphering what the coordinate system is:

pifuhd/apps/recon.py

Line 118 in e47c4d9

color = index(image_tensor[:1], uv).detach().cpu().numpy()[0].T

In summary, I'm trying to use the normal maps predicted by your pretrained pix2pix network, but to do so I need to know how the ground truth normal maps used to train this network were computed.

Thank you!