lsa1997 / POP

Code for "Learning Orthogonal Prototypes for Generalized Few-shot Semantic Segmentation" [CVPR2023]
MIT License
20 stars 3 forks source link

What is the improper setting ? #2

Closed Guzaiwang closed 1 year ago

Guzaiwang commented 1 year ago

As mentioned in your model readme file, "Note that all models are trained based on the released code. We fix an improper setting in the cropping operation and achieve results better than those reported in our paper." I would like to know what is the improper setting and how to fix it.

Thanks

lsa1997 commented 1 year ago

You can check the details here. We just remove the code checking _fg_remainratio. The initial code is used for standard few-shot segmentation, where we check _fg_remainratio to retain enough foreground pixels (_targetcls=1) for binary segmentation. However, it is an improper setting for GFSS since there are multiple foreground classes.