3D Kernel Shape about hausdorffloss HOT 4 CLOSED

Hi, sorry for the long wait. You're right, the shapes for 3D are mismatched. To be honest the erosion version is not really something that works - I've only fiddled with it for a while and haven't really used it or tested it.

from hausdorffloss.

Why do you think it does not work? I had the impression it did work fine.
I rewrote it in plain pytorch.

from hausdorffloss.

I mean, I was under the impression that my implementation is not something that works correctly. As I remember correctly, the gradient propagation was not quite working.

I'm curious, you rewrote this erosion part to be in PyTorch and it seems to work? Did you manage to learn some models on it?

from hausdorffloss.

Yes I did but not as a sole loss, I combined it with a Dice. I got some improvements in the predictions and it did not impaired training performance much after the rewrite. I used it more as an penalty term to modify the Dice score.

from hausdorffloss.