Get final result with new test image

[NguyenHongSon1103]

Hi,
I am confusing about how to determine if an image is anomalous or not because model has two output: classification prediction and GDE scores.
Should I use a threshold for GDE scores to check and ignore model prediction ?
Thank you for your work and your reply :)

[Runinho]

Should I use a threshold for GDE scores to check and ignore model prediction ?

Yes, during inference/validation the paper uses the Density Estimation to detect anomalous data.
The Model is used as a feature extractor on which the density estimation is performed.

During training the whole model, the part that is used as a feature extractor plus a model head, is trained on the cutpast classification proxy task. So the classification output is only used as a proxy task to train the model.

[NguyenHongSon1103]

can you please tell me how to calculate that threshold, i don't see it in your code.

[Runinho]

can you please tell me how to calculate that threshold, i don't see it in your code.

I answered a similar question here:
The thresholding in the evaluation is done by calculating the roc. This might be an option for your implementation (See ROC on wikipedia).

Another Paper by Ripple et al. which Li et. al reference use a probabilistic perspective of calculating thresholds: https://github.com/ORippler/gaussian-ad-mvtec/blob/a2e15800f0087fc7b965e491c53a73f8c7480d1e/src/gaussian/model.py#L152