❓ <title>Inconsistent Heatmap Visualization for Normal Samples During Inference in Dinomaly

[07prince]

Your Question

I’m facing an issue related to heatmap generation during inference when using the Dinomaly model from Anomalib.

When I perform inference on a good (OK) sample, the anomaly score is correctly below 0.5, indicating a normal part.
However, the generated heatmap incorrectly highlights multiple regions as anomalous, even though the sample is good.

Interestingly, when I run inference again on the exact same setup — without changing the camera or part position — but just drop a small piece of paper (defect) inside the bottle,
the anomaly score increases to 1.0, and the heatmap correctly highlights only the defective region.

You can clearly see this behavior in the attached images:

Forum Check

• I have searched the discussions forum for an answer to my question.

[github-actions[bot]]

Thanks for submitting this issue! It has been added to our triage queue. A maintainer will review it shortly.

[rajeshgangireddy]

Hi @07prince
This could be due to the normalisation we do during post-processing step. However, your 'normal' image's map looks like the minimum and maximum values of raw anomaly map generated by Dinomaly are too close.
Try training for fewer more epochs. I noticed that training for more epochs reduces oversensitivity to minor changes in normal images.

if that doesn't work, you might have to try to use a custom post-processor.