1. PatchCore：计算所有feature，再coreset缩减90%，即和训练image的数量N是O(N)的关系。

2. CFA：固定mem bank容量、动态演化，即和N无关。

   1. 记忆库（Memory Bank）+ 度量学习（Metric Learning）

3. Patch descriptor φ(·) is an auxiliary network, which converts pre-trained feature pt into target-oriented features φ(pt).

• Coupled-hypersphere based Feature Adaptation (CFA)
• Modeling of Memory bank C / Compressing mem bank / Updating mem bank
• Test phase

Experiments

CFA++ refers to a case that ensemblesvthe results when using cropped images and using only resized samples.

好像没有比recall，只比较了image和pixel level的precision

1. CFA: Coupled-Hypersphere-Based Feature Adaptation for Target-Oriented Anomaly Localization, ieee access22, 曹吉祥