boundary-loss can not converge

[Xxvn-c]

Hello,

First of all, thank you for your idea. I am encountering some issues when using the boundary-loss:

I have been able to train good models on a custom dataset using ohem-loss and dice-loss. However, I want to add the boundary loss function to enhance the segmentation boundary effects. So I am using “1 * ohem-loss + 1 * dice-loss + 0.01 * boundary-loss”. During training, while ohem-loss and dice-losse converge normally, the boundary-loss does not seem to converge and keeps oscillating around a certain value. Here are some details:

• There are 4 categories for segmentation, and boundary-loss is applied only to one of them.

• The input image resolution is 800x768.

Could you provide some suggestions? Thank you!