Reproducing Fig. 7 with infer.py

[FloorVerhoeven]

Hi!

I'm trying to reproduce the unwarped images in row 3 of figure 7 in the paper but am not managing.
I'm using the infer.py file as provided in this repo, and have both tried with the models available here and here. For pytorch I'm using version 0.4.1. As input images I'm using the 4 images provided in the eval/inp folder.
For example on image 13_2 I get this result

Do you have any ideas what I might need to change in order to be able to reproduce the images in Fig. 7 of the paper?

[sagniklp]

I believe this is happening due to some error in the shape network regression, it could be related to normalization. Could you please share the output of the shape network?

[FloorVerhoeven]

Hi @sagniklp,

Sorry for the delay. This is the output from the shape network (for the 13_2 file):

and as .exr file

[sagniklp]

This issue is resolved if you use scipy (<1.2.0) to load the image. I've updated some instructions in the infer.py accordingly. Following is the wc output you should get if you use scipy:

I haven't done further digging about the difference between OpenCV and scipy imread. I guess it's documented somewhere.

Thanks for bringing this to my attention!

[zxy770624739]

Hello，can you share the code of the output from the shape network？Thank you！