The understanding of loop_cloth iteration in infer.py

[lisunphil]

Thanks for the brilliant and open source work! My issue is that when I read the code of infer.py, two iterations called loop_smpl and loop_cloth confuses me. In my opinion, the loop_smpl iteration is the process that SMPL normal and clothed normal benefits each other. And I think it is the only iteration we need according to the paper. But then there follows loop_cloth. I tried to read the code but still remain confused about the role of loop_cloth iteration in the whole process of infer.py. Before loop_cloth iteration, we have already generated the _recon.obj, and it is definitely with cloth infomation. Hope someone can help me explain what role loop_cloth plays in the code of infer Thanks a lot!

[YuliangXiu]

loop_cloth: refine the reconstructed mesh with predicted normal maps via differential renderer, deform the reconstructed mesh iteratively by minimizing the difference between the normal maps 1) predicted from network 2) rendered from the deformed reconstruction.

This helps a little bit but is quite time-consuming, so I set loop_cloth=0 as default.

[lisunphil]

thx for answering

[YuliangXiu]

@Phil0419 New cloth-refinement module is released. Use -loop_cloth 200 to refine ICON's reconstruction, making it as good as the predicted clothing normal image.