Some unexpected results from your released checkpoints.

[SunnyHaze]

Hi authors!

Thanks for your outstanding work.

However, after running an evaluation script on the CASIAv1 dataset with your official checkpoint trained with CASIAv2, we found that the F1 score is 0.1918. After carefully checking the code I can not find any clues about these unexpected results.

Detailly, I implemented the F1 metrics with the third mask returned from your nn.Modules class.
https://github.com/wangjk666/Objectformer/blob/a967ba91a77e79f6d64e5a02e15f4214197ceb61/ObjectFormer/models/objectformer.py#L379

Here are the visualizing results on CASIAv1:

Do you have any ideas about it? Any help is sincerely appreciated! Thank you very much.