How to match the prediction and targets bounding boxes for MAP

[FrancescoSaverioZuppichini]

Hi Guys,

I hope you are all doing great. I was wondering how to correctly use the new MAP metric. Since targets and predictions have to have the same len, I was wondering how to correctly preprocess the predicted bonding box before passing them to MAP. Since at any step the predicted bounding boxes may be more than the targets, how do I correctly filtered them out? Should I find the matching using iou?

Thanks

[tkupek]

Hi @FrancescoSaverioZuppichini and thanks for the question.

What might be misleading here is the fact that targets and preds is a list of dicts. Each dict represents a sample and contains a list of boxes. While the number of samples in targets and preds has to match, the lists of boxes can vary.

For an example I recommend to have look at this test:
https://github.com/PyTorchLightning/metrics/blob/master/tests/detection/test_map.py

Does this make sense to you?

[FrancescoSaverioZuppichini]

Oh Thanks! Got it ;)