Diffence score F1 between PADIM and me

[davy-blavette]

I have a difference in the calculation of my F1 between what PADIM returns to me and my calculation. I don't know how it comes up with an F1 score of 0.95?

[davy-blavette]

its' ok with threshold = 0.5...

[blaz-r]

This is probably the reason then. Precision and recall depend on the threshold, and anomalib finds the optimal threshold that maximizes the F1 score, then normalizes the predictions in a way that the threshold is at 0.5.

[davy-blavette]

@blaz-r yes, I have a different threfold which I calculate by excluding the most good and keeping the most bad. This gives me a SEA score. I calculate it myself, but perhaps you have something similar?

[samet-akcay]

@davy-blavette, @blaz-r is right. Anomalib returns F1_Max score, based on an adaptive threshold that is computed on the validation set.