Correctness of the adaptive wing loss implementation

[PlumedSerpent]

diff_abs = (target - prediction).abs()
loss = diff_abs.clone()

idx_smaller = diff_abs < theta
idx_bigger = diff_abs >= theta

loss[idx_smaller] = width * torch.log((1 +torch.pow(torch.abs(diff_abs[idx_smaller] / epsilon),(alpha-target[idx_smaller]))))

loss[idx_bigger] = width*(1/(1+torch.pow((theta/epsilon),(alpha-target[idx_bigger]))))*(alpha-target[idx_bigger])*(torch.pow((theta/epsilon),(alpha-target[idx_bigger]-1)))*(1.0/epsilon)*torch.abs(diff_abs[idx_bigger])-\
(theta*(width*(1/(1+torch.pow((theta/epsilon),(alpha-target[idx_bigger]))))*(alpha-target[idx_bigger])*(torch.pow((theta/epsilon),(alpha-target[idx_bigger]-1)))*(1.0/epsilon))-width*(torch.log(1+torch.pow((theta/epsilon),alpha-target[idx_bigger]))))

Hi, I am pretty interested in your excellent paper, and I implemented it in pytorch by myself. However it's wired that I didn't get an obvious improvement compared to MSE Loss. Can you help me find whether my implementation is correct? Or do you have any plan to upload your implementation recently?

[protossw512]

May I ask which task and dataset are you working on and what results did you get for both loss functions?

I ported in your function and get different results compared with mine, your code will have slightly smaller values on pixels where diff_abs >= theta. So there are probably some minor errors in your loss[idx_bigger] part.

For reference, here is my toy input and results:

Prediction: tensor([[1.0000, 0.5000, 0.5000], [0.5000, 0.2500, 0.5000], [0.5000, 0.8000, 0.4000]])

Label: tensor([[0.9800, 0.2500, 0.2500], [0.2500, 1.0000, 0.2500], [0.2500, 0.2500, 0.7500]])

Results: tensor([[0.1740, 1.0378, 1.0378], [1.0378, 8.1260, 1.0378], [1.0378, 4.2064, 3.0384]])

Yours: tensor([[0.1740, 1.0378, 1.0378], [1.0378, 7.8099, 1.0378], [1.0378, 3.9899, 3.0384]])

Hope that helps.

I will release my inference code first during ICCV 2019 (hopefully), and then training code.

[PlumedSerpent]

@protossw512 Thanks for your help. I am working on 300W dataset using HRNet backbone. The result NME is 3.37 with MSE and 3.34 with my adaptive wing loss implementation. And for your toy input, I compute in step by step by hand, it seems my result is correct with theta=0.5,y=0.25,y_hat=0.80, could you please check it again?

[protossw512]

@PlumedSerpent I double checked my code, and you are right, I am on an old experimental branch, which has different settings compared with my paper. I would suggest you try different parameter settings for omega, epsilon, and theta.

[PlumedSerpent]

@protossw512 From your ablation study I don't think different parameter settings would improve a lot. I'll try to re-implement other contributions in your paper to get a better result.