search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.39k stars 652 forks source link

Condinst dice loss #231

Open syusukee opened 4 years ago

syusukee commented 4 years ago

Hi, I've found that in your code you calculated the loss among all instances, not only the positive ones, which you have described in paper. Can you explain it or show me where I can find the related code? Thanks a lot!! @tianzhi0549

tianzhi0549 commented 4 years ago

No, we only compute the mask loss for positive instances, as shown in https://github.com/aim-uofa/AdelaiDet/blob/0157227f966eda93c1299a402537b616207ba226/adet/modeling/condinst/dynamic_mask_head.py#L156. Note than pred_instances only includes the positive ones.

syusukee commented 4 years ago

So is it means FCOS will output the positive proposals which are used in mask branch? @tianzhi0549

tianzhi0549 commented 4 years ago

Yes.

syusukee commented 4 years ago

Ok, I got it. Thanks to your reply!! Really great work!!