1. https://github.com/marco-rudolph/AST
2. 需要先训练一个normalizing flow based teacher from normal images. 这比用大规模数据预训练网络的好？

1. AST[16] concludes that the abnormal image features extracted by the teacher-student model with the same structure are significantly similar, so they propose an asymmetric teacher student architecture to address this issue.

2. AST also introduces a normalized flow to avoid this problem and prevent estimation bias caused by the inconsistency of the two network structures. ???

   

3. tab 5: NF student vs CNN student.

1. teacher: normalizing flow, [22], [38], [39]看来是1d flow，应该用FastFlow的2d flow
   1. FastFlow, 21 对比PatchCore**, cvpr22** 有啥优势？
   2. First, the teacher is optimized to reduce the negative log likelihood loss that may be masked by a foreground map from 3D.
2. student: conventional feed-forward network that does not map injectively or surjectively.
   1. Second, the student is trained to match the teacher outputs by minimizing the (masked) distance between them.
3. mask ⇒ known foreground

Experiments

tab1 Overview of the used datasets.

tab6 inf. time [ms]

References

1. AST: Asymmetric student-teacher networks for industrial anomaly detection, wacv23
2. normalizing flows
   1. [22] CFLOW-AD, WACV22
   2. [38] Same same but differnet: Semi-supervised defect detection with normalizing flows. 22
   3. [39] Fully convolutional cross-scale-flows for image-based defect detection. 22