Anomaly Heatmaps

[SDJustus]

Hi,

is there a way to get the anomaly heatmaps from your code?

Best regards,
Justus

[Pangoraw]

padim.predict(..) returns the mahalanobis distances for each patch of the image. You can resize it to a square matrix to see the unormalized anomaly map. So you can do something like that:

import math
import matplotlib.pyplot
distances = padim.predict(img.unsqueeze(0))
w = int(math.sqrt(distances.numel()))
amap = distances.view(w, w)
plt.imshow(amap)
plt.show()

[Pangoraw]

Here is an example notebook where the input is 416x416 so the backbone outputs 104x104 anomaly maps

[SDJustus]

Thank you again for your super fast work. I appreciate your effort very much!

[rishabh-akridata]

@Pangoraw Why are we using Gaussian blurring as a preprocessing step for the amap? Is it just for visualization or is there any specific reason for doing that?