MeanAveragePrecision for custom area definitions

[chuber1986]

🚀 Feature

MeanAveragePrecision metric computes mAP and mAR for the following area definition:

self.bbox_area_ranges = {
    "all": (float(0**2), float(1e5**2)),
    "small": (float(0**2), float(32**2)),
    "medium": (float(32**2), float(96**2)),
    "large": (float(96**2), float(1e5**2)),
}

I would like to provide a different definition, like:

MeanAveragePrecision(areas={
    "all": (float(0**2), float(1e5**2)),
    "xsmall": (float(0**2), float(16**2)),
    "small": (float(16**2), float(32**2)),
    "medium": (float(32**2), float(96**2)),
    "large": (float(96**2), float(1e5**2)),
})

Motivation

The default does not necessarily make sense for datasets other than COCO.

Pitch

Instead of returning the metrics of the default areas, return the summarized metrics for the provided area definition.

Alternatives

Additional context