What do the various metrics for MAP metric mean?

[dhruvbird]

On https://torchmetrics.readthedocs.io/en/stable/detection/mean_average_precision.html there are a bunch of metrics reported such as map_small, map_medium, map_large. Is there any place where the interpretation of these reported metrics is explained?

Also, is there any additional documentation around the meaning, interpretation, and use of the following options:

1. max_detection_thresholds
2. iou_thresholds
3. rec_thresholds

Thanks!

[dhruvbird]

Okay so it seems like iou_thresholds is a list of IoU values to use when computing mAP. i.e. the AP at each of the values in the list is computed, and then averaged (mean is computed) to produce the mAP metric. I found this information on this page

By default, it seems like the following IoU values are used as thresholds [0.5, 0.55, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95] which is 10 values.

So, if the object detection model is able to predict the bounding boxes but with a lower IoU (say 0.6), it would penalize the mAP metric, since it's averaged across a range of values. The mAP metric would be high if most of the predicted bounding boxes have a high IoU with the ground truth bounding boxes.

If you want to compute the AP metric for all predicted bounding boxes that have an IoU of 0.6 or more, then just pass [0.6] as the argument for iou_thresholds.