search

facebookresearch / MaskFormer

Per-Pixel Classification is Not All You Need for Semantic Segmentation (NeurIPS 2021, spotlight)
Other
1.35k stars 152 forks source link

About ignore value #57

Closed NickChang97 closed 2 years ago

NickChang97 commented 2 years ago

Really thanks for sharing your fantastic work! After browsing your codes, I find it seems that loss of mask predictions has contained the ignore area. No ignore treatment in dice loss and focal loss. Maybe dealing with ignore area in somewhere I do not notice?

bowenc0221 commented 2 years ago

You are right, there is no special treatment to ignore region. Ignore regions are treated as background.