SSIM loss difference

[blaz-r]

Hello.
I have a question regarding the ssim_loss from your library. It seems to produce different results than the two other implementations that I have tried, one from RIAD model and other from DRAEM.

The results of the other two match between each-other, while the kornia ssim_loss is a bit different (even after multiplying by 2).

Following is an example of code:

from kornia.losses import SSIMLoss

draem_loss = SSIM(window_size=11, val_range=1.0) # from DRAEM
riad_loss = SSIMLoss_riad(kernel_size=11) # from RIAD
kornia_loss = SSIMLoss(window_size=11, max_val=1.0)

x = torch.rand(4, 3, 50, 50)
y = torch.rand(4, 3, 50, 50)

print("draem: ", draem_loss(x, y))
print("riad: ", riad_loss(x, y))
print("kornia:", kornia_loss(y, x) * 2)

which produces:

draem:  tensor(0.9139)
riad:  tensor(0.9139)
kornia: tensor(0.9896)

All produce very similar results in training.

What could be the case here if anyone maybe knows?

[edgarriba]

i see that we average in https://github.com/kornia/kornia/blob/master/kornia/losses/ssim.py#L53

[blaz-r]

Hey, thanks for the answer.

I don't see how that line would affect this. Even if I change the line in kornia to loss = 1.0 - ssim_map, the difference persists.

One thing I did notice when comparing the ssim_map produced by either method is that max value differs:
Draem:

ssim_map.max()
tensor(0.8665)
ssim_map.min()
tensor(-0.6337)

Kornia

ssim_map.max()
tensor(0.6580)
>>>ssim_map.min()
tensor(-0.6337)

[edgarriba]

then the only difference is the separable convolution? maybe some of the "border" parameters do not match well

[blaz-r]

Okay, thanks for the answer. I guess these two things could be the case here.