Interpreting OPENVINO results

[LorenzoF6]

Hello, i fit and test on a custom dataset patchcore and padim and the i exported to models to the OPENVINO format. I'm working on Anomalib v2
The i perfom inference using the exported model and some images from my test set.
The test set is composed by two folder : test good images and test defective images

When i perform the inference on test good images the results are something like this: pred.pred_score: 0.993452 pred.pred_label True.

Since i think that the OPENVINO Inferencer say that if a pred_score > 0.5 is considered abnormal and true i supposed is the equivalent of the label ABNORMAL in the previuos versions of the function, my result saying that my model (as patchcore) fail to recognize good test images (because to be a good image the pred_score it has to be < 0.5)

It is correct my reasoning ?
Thank you very much

[FedericoDeBona]

Hi, yes, you are correct. Every model has a threshold set at 0.5 (below is good, above is bad) as they are normalized this way. If you get 0.99 on a test image that should be good, you might have trained the model with incorrect images. Double-check the paths to ensure you are not training with bad images.

[samet-akcay]

Hi guys, I'm converting this to a Q&A, I and believe @FedericoDeBona have answered it. Feel free to continue the discussion here. Thanks