Considering label scores in loss function

[ErezBeyond]

I can't seem to find where in the code is the score of a pseudo box is being taken into account. Specifically, where can we see the effect of zero scored boxes (those that didn't pass the confidence threshold).
To the best of my inquiry, it is missing from the code, though emphasized in the paper.
Thanks!

[zizhaozhang]

Pls see this line

ssl_detection/detection/data.py

Line 232 in 88602f3

mask = pseudo_gt["scores"] >= true_confidence

[ErezBeyond]

Thanks for the reply!
It seem to me that this eventually removes pseudo boxes with lower scores. But this is different than giving them a zero weight, isn't it?
For instance, if the prediction complies with some lower score box, you will penalize for it as a false detection, rather than ignore it (zero weight). Am I missing something here?

[zizhaozhang]

Hi

By design in eq (2) of the paper, we use hard threshold so the w(x) is either 1 or 0 (clearly specified in paper). For 0, we directly remove them.

[ErezBeyond]

Thank you.
For completeness, I would say that a zero weight is not necessarily equivalent to removing the box. It seems that the formulation in the paper is better expressed in setting these boxes as "don't care" areas.