Question about the result

[naeunhub]

I have a question about the model inference result.
The two images below are imaged 7 times and 60 times, respectively, and both images are misclassified. (* image 50 is correct)
When I simply look at the clear image and the reconstructed image, the two images are well reconstructed to normal, but I wonder why the image 7 with clear image clearly defective comes out as normal and the image 60 is abnormal.

I thought it was because of the weight difference between pixel-level distance and feature-level distance, but if you look at image 50, you can see that both pixel and feature distance are high values, but the inference is normal.

Can you explain these results?

woodsample7.png

woodsample60.png

woodsample50.png

[arimousa]

It may root in the normalization in anomaly_map class. Maybe a better design would resolve this problem