mask

[zbc-l]

COTR/COTR/trainers/cotr_trainer.py

Line 143 in 5c9363f

mask = torch.norm(cycle - query, dim=-1) < 10 / constants.MAX_SIZE

why do you use 10 divided by MAX_SIZE?
I can't understand the principle of mask

[zbc-l]

Another strange thing is that when I was registering, the queries were located in the white area of the fixed image, but the points found would be in the black area of the moving image.

[jiangwei221]

The mask is used to enforce cycle-consistency loss. We want to enforce cycle-consistency loss only on already close enough correspondences, not catastrophically wrong correspondences. 10/MAX_SIZE is an empirical number for a good measurement of "close enough".
My best guess for the reasons of the drifting in your case is the textureless white and black regions make the correspondences inaccurate.

[zbc-l]

thanks！
I found a magical thing that cycle_loss calculation, after flipping the image and query, the calculated cycle_loss should be the same as before the flip. Because the essence is to calculate the pred of B from picture A, and then calculate the cycle of A from the pred of B, the difference is only in which A and B are on the left, but the truth is that this two cycle_loss are not the same.