search

xavysp / DexiNed

DexiNed: Dense EXtreme Inception Network for Edge Detection
MIT License
739 stars 157 forks source link

Why use sigmoid-CE balanced loss instead of Dice, Focal, etc? #61

Open za13 opened 3 years ago

za13 commented 3 years ago

Just wondering why Dexined uses sigmoid-CE balanced loss instead of Dice, Focal, etc?

I thought since most pixels in images correspond to the background and not the edges of objects, this would be a class imbalance problem and Dice, Focal, or other losses would handle imbalanced data better than CE?

liuzhy71 commented 3 years ago

I think it has better gradients than Dice.

xavysp commented 3 years ago

Hi, next time I'll consider new loss functions. Now I am facing other equally important problems in edge detection.