IDEA-Research / Stable-DINO

[ICCV 2023] Official implementation of the paper "Detection Transformer with Stable Matching"
https://arxiv.org/abs/2304.04742
Apache License 2.0
218 stars 7 forks source link

Questions about GIoU computed in loss function. #4

Open LMD0311 opened 1 year ago

LMD0311 commented 1 year ago

Thanks for the inspiring work!

But I am confused by the use of not differentiable IoU in the loss function and differentiable GIoU in the matching cost. Or I misunderstanding the methods?

Another question is if there are $C$ classes, $N$ predictions and $M$ GTs, the shape of GIoU during matching will $(N,M)$, but the shape of $pi$ is $(N,C)$, how to get $p{i}f_{2}(s)$ ?