search

JunMa11 / SegWithDistMap

How Distance Transform Maps Boost Segmentation CNNs: An Empirical Study
https://openreview.net/forum?id=hM4pNbXWst
Apache License 2.0
378 stars 67 forks source link

boundary loss based on sdf may be negtive #7

Closed yangyuren03 closed 3 years ago

yangyuren03 commented 4 years ago

HI, Thanks for your great work! I want to know whether the boundary loss may be negtive value. Thanks!

cuicathy commented 4 years ago

I think it is because the distance map within the boundary has negative values.

yangyuren03 commented 4 years ago

@cuicathy ,thanks! Do you have a trial with the boundary loss only and does it convergence?

cuicathy commented 4 years ago

@yangyuren03 With the boundary loss only, the model converges on my dataset, though it is not as good as using the dice loss only.

JunMa11 commented 3 years ago

Hi @yangyuren03 ,

Sorry for my late reply.

I want to know whether the boundary loss may be negative value.

The direct answer is Yes! This is because the sdf takes negative values inside the contour. You can find more details in their paper.