Result of IoU is Nan when there is empty item in preds.

[joshuasv]

🐛 Bug

To Reproduce

The target is:

[
 {'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]]), 'labels': tensor([0, 0, 0])}
]

The preds is:

[
{'boxes': empty(size=(0, 4)), 'labels': tensor([], dtype=torch.int64), 'scores': tensor([])}
]

The last item of preds is empty, when using torchmetrics.detection.iou.IntersectionOverUnion to calculate IoU, the result is nan.

Code sample

import torch
from torch import tensor
from torch import empty
from torchmetrics.detection import IntersectionOverUnion

def calc_iou(preds: list, target: list):
    metric = IntersectionOverUnion()
    metric.update(preds, target)
    result = metric.compute()
    return result

target = [
 {'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]]), 'labels': tensor([0, 0, 0])}
]
preds = [
{'boxes': empty(size=(0, 4)), 'labels': tensor([], dtype=torch.int64), 'scores': tensor([])}
]

calc_iou(preds, target)

{'iou': tensor(nan)}

Expected behavior

A correct iou value, i.e: 0.0

Environment

• TorchMetrics version (if build from source, add commit SHA): torchmetrics==1.4.0.post0
• Python & PyTorch Version (e.g., 1.0): Python 3.11.9 & torch==2.3.1+cpu
• Any other relevant information such as OS (e.g., Linux): Ubuntu 22.04.5 LTS

Additional context

[github-actions]

Hi! thanks for your contribution!, great first issue!