What is the ground truth for training?

[rginpan]

Thanks for your work. When training the focal loss and multi-category score predictor, do you need the foreground and background label (ground truth)

[linxid]

We use COCO label to generate foreground and background label. Pytorch code is avaliable in https://github.com/IDEA-Research/detrex/blob/main/projects/focus_detr/modeling/two_stage_criterion.py

[rginpan]

Thanks for your reply, about multi-category score predictor, does it mean select the top-k score in foreground token? "multi-category" is a little bit confused me.

[linxid]

"Multi-category score predictor " is similar to topk token selector in the Encoder of DINO.