Width Calculation For Bounding Boxes in torchvision\models\detection\_utils.py

[JamesMcCullochDickens]

In the function encode_boxes (line 79 of torchvision\models\detection_utils.py), it seems that the width of the ground truth proposals matched is being computed as

ex_widths = proposals_x2 - proposals_x1
ex_heights = proposals_y2 - proposals_y1

But for a bounding box from ms coco [368, 413, 368, 417]. I guess this is just a matter of opinion if this is a "valid" bounding box, but it seems to me that x_min = x_max is valid for a box that is 1 pixel wide, and y_max-y_min pixels high. Anyway this causes the targets_dw or targets_dh to take the torch.log of 0, giving float(-inf), which can of course be easily fixed by adding +1 to the width, or the fix:

ex_widths = proposals_x2 - proposals_x1 + 1
ex_heights = proposals_y2 - proposals_y1 + 1

Either that or I could just filter out these boxes with x_min = x_max or y_min = y_max

[fmassa]

Hi,

Thanks for opening this issue. We have discussed about the decision of dropping the +1 in #1872, and more specifically in #826 (comment)

I'm closing this issue as a duplicate, but let us know if you have further questions