Enable AUPRO (Per Region Overlap) metric to support mask

[triet1102]

Hello, I am currently using AUPRO to evaluate anomaly localization performance and it seems to be that we are not able to filter out points when calculate the metric?

I am asking if we can add support for masking to filter out unnecessary points?
A workaround that I am doing is

• create a wrapper class for AUPRO
• override the init class which adds inspection_mask parameter
• filter out unnecessary points right after the connected component analysis

  anomalib/src/anomalib/utils/metrics/aupro.py

  Lines 87 to 94 in c92fb52

  	if target.is_cuda:
  	cca = connected_components_gpu(target)
  	else:
  	cca = connected_components_cpu(target)
  	preds = preds.flatten()
  	cca = cca.flatten()
  	target = target.flatten()

Thanks in advance

[ORippler]

I guess you could split/refactor the _compute function into two separate perform_cca and compute_pro functions. This would give you the ability to modify preds/cca/target based on your masking without forcing the need to support and test this in anomalib.

[triet1102]

OK thanks @ORippler , I'll wait for the verdict before proposing a PR for that.

[ORippler]

CC @samet-akcay before opening a PR

[samet-akcay]

Thanks for the reminder @ORippler! I've missed this.

Splitting _compute to perform_cca and compute_pro functions sounds good to me.

[samet-akcay]

Closing this as it has already been implemented.