Abnormal segmentation prediction without pred_mask

[pirnerjonas]

Hi,

I just encountered a situation where I predict abnormal for an image but pred_mask doesn't show anything (image_result.pred_mask.sum() is 0):

is this to be expected?

[yxteo2]

Can show how you get the predicted mask?

[alexriedel1]

Which type of model are you using?

what you see probably happens because the threshold for an whole Image to be anomalous and the treshold for a pixel on the anomaly map to be anomalous are calculated separately.
You can have a deeper look into these thresholds to find out if that is the case

[pirnerjonas]

I think for that test it was Padim. And yes I think it was like you said @alexriedel1. Image threshold influences pred label, pixel threshold influences pred mask.