IOU type for mean average precision

[gianscarpe]

🚀 Feature

would be nice to have detection.map.MeanAveragePrecision similar to what pycocotools provides. In particular, introduced a parameter "iou_type" in order to choose between "bbox" and "segm" (for instance segmentation).

Motivation

Having a metric for instance segmentation, which is currently unavailable.

Pitch

Instantiate MeanAveragePrecision(..,, iou_type="segm") to evaluate instance segmentation

Alternatives

Additional context

[github-actions]

Hi! thanks for your contribution!, great first issue!

[drozzy]

Could you explain the difference between bbox iou and the segmentation IoU?
I saw a reference to this in the torchmetrics docs and wasn't sure what that meant.

[Borda]

@gianscarpe ^^

[gianscarpe]

Could you explain the difference between bbox iou and the segmentation IoU? I saw a reference to this in the torchmetrics docs and wasn't sure what that meant.

Hi and thanks for the question! As in pycocotools (the official COCO library for mAP), the difference between the two regards how to compute positive and negative pairs of bounding-box. BBox_IOU computes how much the bounding-boxes overlap, while segmentation-iou evaluates how much the segmentation masks (ground-truth and predicted). Different way to compute intersections and unions :)

[drozzy]

Is it perhaps similar to the Non-Max Suppression? Pages 41-45:
https://web.eecs.umich.edu/~justincj/slides/eecs498/FA2020/598_FA2020_lecture15.pdf

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[gianscarpe]

PR: #822