Strengthening the Normalizing Flows by Contrastive Learning for Better Anomaly Detection

Abstract

说明有作用，推开了

说明，不简单，abnormal的构造重要。

Introduction

1. 2d NF出正负样本的Z和Neg_Z，再接contrastive learning，

2. smooth classification？

   

   1. where CELoss represents CrossEntropyLoss, and τ is the label smoothing coefficient.

3. contrastive learning: 只考虑了正样本的全局特征要靠近，正负样本的box内的特征要排斥。

   1. Z⇒S

   2. Neg_Box_s represents the features of anomalous regions in negative samples

   3. Box_s represents the features of corresponding regions in positive samples.

   4. F denotes the cosine similarity function

      

retaining the lightweight characteristics of the 2D-Flow 为啥2d更light？

Related work

Normalizing Flow

1. BGAD-FAS [37]
   1. initially used self-generated negative samples to optimize classification boundaries during training, the two-stage process and manual boundary delineation posed inconvenience.

References

1. CL-Flow: Strengthening the Normalizing Flows by Contrastive Learning for Better Anomaly Detection, 23
2. BGAD-FAS [37]: Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection (CVPR2023)