search

wogur110 / PNI_anomaly_detection

Official code for the paper "PNI : Industrial Anomaly Detection using Position and Neighborhood Information"
33 stars 8 forks source link

Position contribute #1

Closed FilippoComastri closed 2 months ago

FilippoComastri commented 11 months ago

Hi. When you compute histograms relative to position information ( p(c|Cdist)), why you use embedding coreset (Cemb) instead of distribution coreset (Cdist) , as explained in the paper? As I see you used Cdist only for neighourhood contribute but not for position. I'm interested to know if there's a specific reason for this choice.

wogur110 commented 11 months ago

Hi. distribution coreset (Cdist) is subset of embedding coreset (Cemb). (|Cdist| < |Cemb|) I use embedding coreset (Cemb) to compute histogram because it can output more accurate normal feature distribution (p(c|Cemb). On the other hand, I use dsitribution coreset (Cdist) to compute neighborhood contribute due to limited size of MLP.