mAP - Ignoring false positives?

[ChristofferEdlund]

Hey.

When exploring torch metrics mAP we noticed that passing a GT with one box and a prediction with multiple boxes the mAP metric seem to ignore the false positives when calculating precision.

Is this the sought behavior of the mAP metric or am I employing it wrong?
(See image for example)

Code to reproduce:

from torchmetrics.detection.mean_ap import MeanAveragePrecision
import torch

map_metric = MeanAveragePrecision()

tm_gt = [{"boxes" :torch.tensor([[30, 30, 40, 40]]), "labels": torch.tensor([1]), "scores": torch.tensor([1])}]
tm_preds = [{"boxes": torch.tensor([[30, 30, 40, 40], [30, 25, 40, 35], [10, 10, 20, 20]]), "scores":torch.tensor([.8, .8, .8]), "labels":torch.tensor([1, 1, 2])}]


map_metric.forward(tm_preds, tm_gt)

output:

{'map': tensor(1.),
 'map_50': tensor(1.),
 'map_75': tensor(1.),
 'map_small': tensor(1.),
 'map_medium': tensor(-1.),
 'map_large': tensor(-1.),
 'mar_1': tensor(1.),
 'mar_10': tensor(1.),
 'mar_100': tensor(1.),
 'mar_small': tensor(1.),
 'mar_medium': tensor(-1.),
 'mar_large': tensor(-1.),
 'map_per_class': tensor(-1.),
 'mar_100_per_class': tensor(-1.)}