Bugfix for empty preds or target in iou scores

[SkafteNicki]

What does this PR do?

Fixes #2805 When either preds or target have empty boxes the score for that pair should be 0. Currently instead it is a empty tensor which then are not counted towards the global average making the scores too high.

Before submitting

- [ ] Was this **discussed/agreed** via a Github issue? (no need for typos and docs improvements) - [ ] Did you read the [contributor guideline](https://github.com/Lightning-AI/torchmetrics/blob/master/.github/CONTRIBUTING.md), Pull Request section? - [ ] Did you make sure to **update the docs**? - [ ] Did you write any new **necessary tests**?

PR review

Anyone in the community is free to review the PR once the tests have passed. If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

---

📚 Documentation preview 📚: https://torchmetrics--2806.org.readthedocs.build/en/2806/

[codecov[bot]]

Codecov Report

Attention: Patch coverage is 60.86957% with 9 lines in your changes missing coverage. Please review.

Project coverage is 35%. Comparing base (4c75369) to head (2c2993c). Report is 3 commits behind head on master.

:exclamation: There is a different number of reports uploaded between BASE (4c75369) and HEAD (2c2993c). Click for more details.

HEAD has 296 uploads less than BASE

| Flag | BASE (4c75369) | HEAD (2c2993c) | |------|------|------| |Linux|59|9| |torch2.0.1+cpu|18|3| |python3.8|6|1| |cpu|88|14| |macOS|17|3| |python3.12|18|3| |torch2.5.0|6|1| |python3.10|57|9| |Windows|12|2| |torch2.5.0+cpu|6|1| |torch2.0.1|11|2| |python3.11|7|1| |torch2.4.1+cu121|15|2| |torch2.2.2+cpu|6|1| |torch2.1.2+cpu|6|1| |torch2.3.1+cpu|6|1| |torch2.5.0+cu124|14|2|

Additional details and impacted files

```diff @@ Coverage Diff @@ ## master #2806 +/- ## ======================================== - Coverage 69% 35% -34% ======================================== Files 338 338 Lines 18334 18352 +18 ======================================== - Hits 12586 6387 -6199 - Misses 5748 11965 +6217 ```

[yurithefury]

@SkafteNicki thanks for implementing the fix! I just noticed another potential unexpected behaviour is when both targets and preds contain empty boxes:

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

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

iou = IntersectionOverUnion()
iou.update(preds, targets)
result = iou.compute()  

This yields an IoU of 0, but it probably should return 1

[SkafteNicki]

@SkafteNicki thanks for implementing the fix! I just noticed another potential unexpected behaviour is when both targets and preds contain empty boxes:

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

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

iou = IntersectionOverUnion()
iou.update(preds, targets)
result = iou.compute()  

This yields an IoU of 0, but it probably should return 1

@yurithefury I do partly not agree. When both target and preds contains empty boxes both the intersection and union of those boxes are 0, meaning that we end up trying to compute 0/0, which is undefined. It would maybe make the most sense setting it to Nan, but that would mess with average over multiple samples where only one is not defined. So there is really no right or wrong value in this case. Maybe it needs to be an argument to control?

[yurithefury]

@SkafteNicki Good point! Allowing an option to set this value, like empty_iou_value, would give flexibility - for example, setting it to NaN, 0, 1, or just ignoring these cases altogether. This way, users could pick what fits their use case best.