Open zhangchbin opened 5 months ago
I have the same confusion. In this code, the selection criteria for the first (num_queries=300) objects seem to be just classification scores, rather than combining classification and localization as the author said.
hello, do you find the related code?
Hi, I think it is this part https://github.com/lyuwenyu/RT-DETR/blob/14971513a5ef805edaf434af1a11403d2f1a9036/rtdetr_pytorch/src/zoo/rtdetr/rtdetr_criterion.py#L111. The authors mentioned they used it for the classification loss
Thanks to the authors' work. I can not find the uncertainty minimal query selection module in the pytorch code. It uses a classification score to select tokens, as shown in https://github.com/lyuwenyu/RT-DETR/blob/main/rtdetr_pytorch/src/zoo/rtdetr/rtdetr_decoder.py#L488C1-L489C1