1. APG-Net改进了MulSen-AD, cvpr25 的memory bank via a non-parametric feature space reshaping paradigm: Prototype matching + adaptive prototype guidance
   1. 认为直接预训练的feature辨识度不够，所以要让正常特征的分布进一步收缩，已增加和异常特征的距离。

• Prototype matching

1. Adaptive prototype guidance: *APG

   1. To enhance the separability between normal and abnormal features

   2. 每个feature，向其对应的P*靠拢，即分布收缩

      1. stage 1 计算收缩权重，距离P*近，则权重大；c is a scaling factor, and τ(·) represents the Sigmoid function.
      2. stage 2 向P*平移

   3. Testing Phase

      If the test features are abnormal, they already deviate from the normal distribution. Their norms are large, and the corresponding α values are small.

Discussion

• Why EMA? APG-Net为啥不提前统计好全部normal feature的mean和variance，再做feature normalization，build memory bank；而采用分batch，用EMA的方式来做？
• What is EMA

References

1. APG-Net: Adaptive Prototype Guidance Network for Multi-Sensor Industrial Anomaly Detection, tcsvt26
2. Multi-Sensor Object Anomaly Detection, Unifying Appearance, Geometry, and Internal Properties, cvpr25