search

wuhuikai / FastFCN

FastFCN: Rethinking Dilated Convolution in the Backbone for Semantic Segmentation.
http://wuhuikai.me/FastFCNProject
Other
838 stars 148 forks source link

Questions about the SE-loss and Aux-loss #64

Closed meanmee closed 4 years ago

meanmee commented 4 years ago

Hi, first thank you for the great work. I just checked the codes and also had run some scripts. I am confused with the final loss which is composited with three individual losses. could you tell what is the se-loss and the aux-loss used for.

wuhuikai commented 4 years ago

Please refer to Context Encoding for Semantic Segmentation. In a word, se-loss is sth like classification loss and aux-loss is deep supervision.

meanmee commented 4 years ago

Is se-loss good at solving data-unbalanced problems? cuz my experiments based on my own dataset showed this benefit. actually I don't know if if is because of se-loss. the results are just getting better trained on your codes

wuhuikai commented 4 years ago

I think that se-loss can solve aera-unbalanced problems in a degree.

meanmee commented 4 years ago

image Why the results are like this. the while lane should be in a solid way instead of the dot form. Is it because of the dilation convolution?

wuhuikai commented 4 years ago

I'm not sure. Maybe you can try dice loss.

meanmee commented 4 years ago

But it works fine when training it with BiSeNet with just one single ce-loss.

wuhuikai commented 4 years ago

Maybe you can add a 3x3 conv at the end of JPU.