The result is Nan when computing IoU in objection detection.

[fansuregrin]

The target is:

[
 {'boxes': tensor([[ 86.0000,  15.0000, 214.0000, 224.0000],
                            [201.0000, 109.0000, 294.0000, 141.0000]], device='cuda:2'), 'labels': tensor([ 6, 10], device='cuda:2')}, 
 {'boxes': tensor([[124.,  79., 399., 237.],
                            [  0.,  82.,  66., 156.]], device='cuda:2'), 'labels': tensor([5, 6], device='cuda:2')}, 
 {'boxes': tensor([[238.0000,  74.0000, 343.0000, 275.0000]], device='cuda:2'), 'labels': tensor([6], device='cuda:2')}, 
 {'boxes': tensor([[  8.0000,  70.0000,  76.0000, 110.0000],
                            [247.0000, 131.0000, 315.0000, 175.0000],
                            [361.0000, 177.0000, 395.0000, 203.0000]], device='cuda:2'), 'labels': tensor([0, 0, 0], device='cuda:2')}
]

The preds is:

[
{'boxes': tensor([[ 69.0789,  32.9795, 237.3270, 223.9482],
        [192.3514, 108.7430, 293.3430, 144.3972],
        [194.6256, 108.4384, 291.3083, 143.7163],
        [195.6599, 113.8255, 291.3373, 144.4118]], device='cuda:2'), 'labels': tensor([ 6,  5, 10, 13], device='cuda:2'), 'scores': tensor([0.9997, 0.8557, 0.6551, 0.1529], device='cuda:2')}, 
{'boxes': tensor([[116.1356,  70.3358, 353.5793, 214.7886],
        [  7.3007,  98.0802,  69.1378, 163.0365],
        [305.5327,  27.2811, 322.0698,  39.8957],
        [112.4302, 208.0613, 173.9218, 292.1096]], device='cuda:2'), 'labels': tensor([5, 6, 5, 6], device='cuda:2'), 'scores': tensor([1.0000, 0.9450, 0.5993, 0.1131], device='cuda:2')}, 
{'boxes': tensor([[222.7714,  74.0670, 365.3134, 249.2353],
        [201.2384,   4.7298, 230.6123,  16.7820],
        [322.8268, 151.7757, 400.0000, 187.3730],
        [140.8015, 170.1787, 227.3417, 291.6501],
        [100.7026,  64.1443, 367.1241, 231.8164],
        [ 17.2630,  38.2540,  74.0354, 209.2324],
        [  0.0000, 155.9979,  78.3138, 208.8021],
        [  3.1740, 149.3114,  73.7855, 210.7626]], device='cuda:2'), 'labels': tensor([ 6,  5,  5,  6,  6,  6, 13,  6], device='cuda:2'), 'scores': tensor([0.9951, 0.9907, 0.9418, 0.8937, 0.7300, 0.2960, 0.1127, 0.1039],
       device='cuda:2')}, 
{'boxes': tensor([], device='cuda:2', size=(0, 4)), 'labels': tensor([], device='cuda:2', dtype=torch.int64), 'scores': tensor([], device='cuda:2')}]

The last box of preds is empty, when I using torchmetrics.detection.iou.IntersectionOverUnion to calculate IoU, the result is nan. So, I want to ask if I remove the empty box from preds, and then re-compute the iou. Is this iou the correct one?

[fansuregrin]

I have found merely removed the empty item in preds will raise a ValueError. In other words, the preds and target must have same length. So, how to make this calculation correctly? Could somebody give me a hint. Thx a lot.