Anomalib image_threshold with recall=1

[mohamed-s7]

in anomalib PaDiM, model.image_threshold = 88.8 for the best F1 score.
but I want the recall = 1 even if I lose a little bit in precision, so I want recall = 1 and maximise the precision.
how can I find that threshold?

using just pred_scores that are between 0 and 1, and labels the threshold that I am getting is 0.4 (value between 0 and 1) but manually I found the threshold that makes the recall = 1 and maximaizes the precision is 77.8

[alfieroddan]

Each model in anomalib generates a prediction by using the PostProcessor.

So in your case, instead of calculating the threshold based on F1 (this is done internally by using F1AdaptiveThreshold), you could:

Turn off post-processing

from anomalib.models import Padim

model = Padim(post_processor=False)

This would generate an "anomaly map" which you can do any custom normalisation, thresholding or processing you want.

Change PostProcessor Args

The PostProcessor has some args you can change.

As an example you could keep the normalisation of the PostProcessor but turn off the thresholding.

from anomalib.models import Padim
from anomalib.post_processing import PostProcessor

post_processor = PostProcessor(enable_thresholding=False)
model = Padim(post_processor=post_processor)

Then you could manually set the threshold but keep the normalisation that anomalib does.

Create your own PostProcessor

See here in the docs

Implement a new adaptive threshold

I think this is probably the best way for longevity is to create your own PostProcessor with your own threshold selection. Internally for PaDiM the default PostProcessor is used.

Instead of F1AdaptiveThreshold, you could write your own!

This could adaptively set the threshold for optimal Recall whilst maintaining anomalibs PostProcessor features.

[samet-akcay]

Thanks @alfieroddan for the detailed response. I'm converting this to a Q&A. We could continue there if needed