Handling negative images or images with no object of interest

[PrajwalCogniac]

The images I work with are damage to industrial parts with a very specific image, for that reason a defective part and another one that is fine are very similar.

The problem of taking samples randomly, is that although it is true that many parts of the background are taken there are practically no good samples of parts without defects. What ends up happening is that it relates the defective being with the piece itself (since it has not seen enough).

That is why I want to give feedback to the system so that it sees completely clean images, without objects.

I will try to make the necessary changes in the MultiBox_Loss, I think I should only add the check if the target is empty and avoid the match. Line 105

And then add checks on those elements that depend on the anchors (localization and mask). Probably return a 0 directly?

Originally posted by @adriaciurana in #137 (comment)
@adriaciurana I am facing the exact issue, I am doing on industrial images as well. Can you please let me know how did you handle this ?

Any help on this issue is useful @adriaciurana @dbolya @abhigoku10