[Question] Model Misclassified Normal Images

[thodung003]

Hello everyone, I have just tried using Patchcore for training custom data.
The model's error detection is impressive, but in real-world usage, the heatmap score for normal products is very noisy, and no threshold can effectively distinguish between normal and abnormal images.

real abnormal

normal

How do you usually handle such cases to solve the issue of false positives? Thank you, everyone.

[abc-125]

Hello, if the test metrics belong to your custom data, you should be able to find the threshold for the image-level classification task.

The pixel-level segmentation (visualised on the heatmap) is another problem. Heatmaps are usually quite noisy for PatchCore; you can try to set a higher minimum value for visualisation to have less noise.

[thodung003]

Hello, if the test metrics belong to your custom data, you should be able to find the threshold for the image-level classification task.

The pixel-level segmentation (visualised on the heatmap) is another problem. Heatmaps are usually quite noisy for PatchCore; you can try to set a higher minimum value for visualisation to have less noise.

Thank you for your support. My custom data is quite similar to Mvtec zipper, and I have tried it with the zipper dataset.

In that case, there were also instances of false errors. If I increase the higher minimum value for visualization, there is a risk of missing real errors, as the score for real errors often matches the noise in many cases.

[abc-125]

Unfortunately, there is always a trade-off between detecting real errors and avoiding false positives unless you have image_AUROC 1.0 (perfect separation between classes).

[thodung003]

Thanks for your support