search

YBIGTA / pytorch-hair-segmentation

pytorch-hair-segmentation
MIT License
457 stars 90 forks source link

HairMattingLoss Problem #33

Closed tchaton closed 5 years ago

tchaton commented 5 years ago

Hello there,

I would replace torch.sum(torch.mul(G, 1 - torch.pow(I_xG_x + I_yG_y,2)))/torch.sum(G) + 1e-6 by torch.sum(torch.mul(G, torch.abs(1 - torch.pow(I_xG_x + I_yG_y,2))))/torch.sum(G) + 1e-6.

I tried it with only this loss, and it was negative.

Best, T.C

davinnovation commented 5 years ago

Thx for interest our project. Our HairMattingLoss for calculating "mask-image gradient consistency loss" part is quite buggy, so we didn't use it temporarily. Add option for your suggestion in this PR

If we use abs for that term, it can explode, so we add range [0,1]. If there's any other better solution, PR is always welcome. Thank you

arjitkatare commented 4 years ago

One need to ensure that square root and division are not getting nan by adding some small number like 1e-6 as mentioned in this code, add that in square root part too. images were more refined after that as mentioned in the paper, though for me there were still some extra edges from other part of the picture were appearing, they were little none the less. Good enough for hair colour change matting process as mentioned in the original paper.