What is the IOU interval for AP for bbox eval?

[pjoyce1995]

I am referring to the AP listed in the table broken down by category.

I cannot find the appropriate directory to find this myself. It seems that the mean_ap.py code in mmdet.core.evaluation isn't used when running ./tools/test.py

The mean_ap.py calculates AP by the '11 points' as the average between IOU of 0 to 1, but other sources say that COCO uses IOU between 0.5 and 0.95.

[zachcoleman]

Most evaluation is driven by the dataset you are evaluating against and in the corresponding code. For COCO you are correct:

mmdetection/mmdet/datasets/coco.py

Lines 616 to 620 in 9d3e162

	iou_thrs (Sequence[float], optional): IoU threshold used for
	evaluating recalls/mAPs. If set to a list, the average of all
	IoUs will also be computed. If not specified, [0.50, 0.55,
	0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95] will be used.
	Default: None.

For the mmdet CustomDataset it is here:

mmdetection/mmdet/datasets/custom.py

Line 327 in 9d3e162

iou_thr (float | list[float]): IoU threshold. Default: 0.5.

It is whatever you configure with your test dataset for the evaluate function to use.