Questions

PatchCore-10, PatchCore-25什么意思？

Introduction

  1. PatchCore: **Towards Total Recall in Industrial Anomaly Detection, cvpr22**
    1. https://github.com/amazon-science/patchcore-inspection
    2. most famous prototype-based method
    3. Other
      1. Ranked #3 on Anomaly Detection on AeBAD-V
      2. Ranked #9 on MVTec AD
      3. Ranked #8 on VisA
      4. 只是在MVTec AD的结果上，Total Recall高，并没阐述本方法为啥recall高。

goal

  1. full recall: fig s4
  2. few-shot: table s5 and fig 6.

Method

Untitled

uses a network φ pre-trained on ImageNet.

φi,j = φj (xi) to denote the features for image xi ∈ X (with dataset X ) and hierarchy-level j of the pretrained network φ.

j ∈ {1, 2, 3, 4} indicating the final output of respective spatial resolution blocks.

Locally aware patch features

patch size p = 3;

Untitled

  1. use features of different levels instead of the last level of a backbone
    1. use features comprising intermediate or mid-level feature representation, otherwise ⇒
      1. loses more localized nominal information;
      2. no j==1: too local
      3. do not use j==4: very deep and abstract features in ImageNet pretrained networks are biased towards the task of natural image classification, which has only little overlap with the cold-start industrial anomaly detection task
  2. Locally aware patch features
    1. locally aware: average pooling of patch features around current position as new patch features of current position
      1. the feature is from some mid-level, so it is actually a patch feature instead of a pixel feature.
  3. how to use multiple feature hierarchies / levels?
    1. rescaling, them element aggregation
  4. locally aware patch-feature collection Ps,p(φi,j )

image-level memory bank of patch features

Untitled