Implementation not consistent with the original paper?

[Joker316701882]

Hi.
In original paper, it mentioned in Sec. 4 that

To optimize for AP, we override the prediction of these slots with the second highest scoring class, using the corresponding condence. This improves AP by 2 points compared to filtering out empty slots.

But I didn't see any corresponding code in this repo. Did I miss something or it is not implement here?

Thank you.

[alcinos]

Hello,
It is implemented in the post-processor, namely here:

detr/models/detr.py

Lines 265 to 266 in b7b62c0

	prob = F.softmax(out_logits, -1)
	scores, labels = prob[..., :-1].max(-1)

As you can see, the softmax is taken over all the classes, but the max itself excludes the last element (no-object).

I believe I have answered your question, as such I'm closing this. Let me know if something is still unclear.

[Joker316701882]

@alcinos
This is exactly what I was looking for.
Thank you very much.