Generalized IOU loss doesn't handle 0-length and -width boxes correctly, expected 0.0 but got 1.0

🐛 Describe the bug

Generalized IOU loss doesn't handle boxes that are parametrized as 0 length and 0 width correctly. The expectation is that with identical boxes, the GIOU loss is close to zero. This is true for non-trivial boxes with positive length and width. For trivial boxes [0, 0, 0, 0] compared against itself, the GIOU loss returns 1. where we expected the loss function to return 0. This is an edge case.

In the implementation, this is because iouk is 0 when iou=0 and union=0 due to the epsilon in the denominator. Can we catch situations where iou == union == 0, and return 1?

Hoping this behavior can be caught and aligned, so that GIOU returns 0 for trivial boxes too.

>>> boxes = torch.tensor([[0, 0, 0, 0]])
>>> torchvision.ops.generalized_box_iou_loss(boxes, boxes)
tensor([1.])  # this should be 0
>>> boxes = torch.tensor([[0, 0, 1, 1]])
>>> torchvision.ops.generalized_box_iou_loss(boxes, boxes)
tensor([1.1921e-07])

Versions

most recent version

Start with the generalized_box_iou_loss implementation linked in the issue and inspect how zero-area boxes affect IoU, union, and the final loss. Reproduce the two examples, then add or update coverage for identical zero-length, zero-width boxes and verify that the loss is 0 while existing positive-area behavior remains unchanged.