About the seismic colormap

[suger111]

Hi，Can the code to get the seismic colormap be made publicly available？

[suger111]

Or how do you view the test results of a test set for a specific image？

[ORippler]

Or how do you view the test results of a test set for a specific image？

Our method does not provide anomaly segmentation results per-se, as we aggregate over all spatial locations by means of average pooling before modeling the PDF by means of MVG. If you are interested in anomaly segmentation results, I refer you to PaDiM, which extended our approach to model the PDF "per-pixel"/spatial location inside the intermediate feature representations by means of MVG. A good implementation for PaDiM is contained in anomalib, and we strongly recommend anomalib in general, as you can find many other algorithms also there.

Can the code to get the seismic colormap be made publicly available？

The heatmaps in our paper are generated by applying XAI techniques, i.e. they are a means to attribute how much the individual pixels contributed to the image-level anomaly score by means of modified backpropagation. Such techniques are in general more compute intensive than those that generate segmentations directly (e.g. PaDiM/PatchCore), and suffer from their own drawbacks (refer also the paper). If you are interested in XAI, I can try to make the time to publish the code also here (iirc we were refactoring it but never got around to publishing it). Here, I would recommend you to apply XAI to the heatmaps generated by PaDiM/PatchCore to resolve the non-linear contribution of the input pixels to the per-location anomaly scores as a direction of future research (this would require numerous code changes though).

Hope this helps.

[suger111]

OK. Thank you so much.