Fix infinity in target boxes

[lemairecarl]

I got -inf values in my target boxes, which propagated up to the loss. I tracked down this line in encode() where a log is taken on the width/height of the boxes. For some reason, some of my data-augmented boxes have zero-width (or height).

It probably happens in the data augmentation, when cropping, with boxes that are on the side of the image. One solution would be to look for zero-width/height boxes and remove them.

But I think it's a good practice to completely remove the risk of -inf by shifting the logarithm a tiny bit, like I do in my suggested change.

What do you think?

[rronan]

I confirm I faced the same issue and fixed it the same way. Still don't know what is a good value for the "tiny shift" (@lemairecarl suggests 1e-10).

[isthatyoung]

Thank you for your inspiration!!! My dataset has just 5 "dirty" data with zero width/height box, now my "inf" problem was solved, thanks.