GIoU Loss becomes NAN

[godprincelp]

The GIoU loss becomes NaN when the gt boxes and predicted boxes are both decoded in center and size (xc, yc, w, h).

I am wondering if the boxes have to decoded in two corners (x1, y1, x2, y2).

[weichen456]

I have encountered the same problem

[shrezatofighi]

The GIoU loss becomes NaN when the gt boxes and predicted boxes are both decoded in center and size (xc, yc, w, h).

I am wondering if the boxes have to decoded in two corners (x1, y1, x2, y2).

No they don't have to. NaN may occurs when the loss stability cannot be ensured, e.g. when the outputed values for w or h can be non-positive.

[sbastani]

The same problem here. Anyone has managed to solve it?

[notabee]

same problem