[paper notes] catching both gray and black swans: open set supervised analog detection*

The paper

Thesis title ：Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection*

Included ：CVPR2022

Address of thesis ：[2203.14506] Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection (arxiv.org)

Project address ：GitHub - Choubo/DRA: Official PyTorch implementation of the paper “Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection”, open-set anomaly detection, few-shot anomaly detection.

Thesis translation ：Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection_appron The blog of -CSDN Blog

This paper is about the problem of image anomaly detection on open data sets , In this paper, a lot of knowledge frame information is hidden , If you want to know more about it, you can read another paper first 《Deep Anomaly Detection with Deviation Networks》, Several interpretations are very clear , After reading it, you can basically understand the framework of this article .

《Deep Anomaly Detection with Deviation Networks》

Source of the paper ｜ KDD 2019
Thesis link ｜ [1911.08623] Deep Anomaly Detection with Deviation Networks (arxiv.org)
Source link ｜ GitHub - GuansongPang/deviation-network: Source code of the KDD19 paper "Deep anomaly detection with deviation networks", weakly/partially supervised anomaly detection, few-shot anomaly detection

Interpretation of the thesis | DevNet： Deep anomaly detection model based on deviation network | Dreamhouse blog (dreamhomes.top)

Paper sharing | Deep Anomaly Detection with Deviation Networks (qq.com)

DevNet Semi supervised anomaly identification model - You know (zhihu.com)

Interpretation of the overall framework diagram ：Deep Anomaly Detection with Deviation Networks

 

Interpretation of the thesis Catching Both Gray and Black Swans: Open-set Supervised Anomaly Detection*

Data set presentation  

  

Paper improvement

primary coverage & contribution

  Key points

【 Reading papers 04】CVPR2022 the selected readings _ Be humble about people's blogs -CSDN Blog

Problem definition

  Ideas

frame

Code reading

4 individual head 

  Data preprocessing

Visible abnormal sample reading  

  Pseudo exception sample generation

  Model input

Decoupling anomaly score  

Return value

Loss 

​

BEC Loss、Focal Loss and Dev Loss Comparison  

Separate abnormal learning , various head Of loss 

experiment

Details of the experiment  

Normal setting result

Difficult setting results

Ablation Experiment

Every anomaly learns head Importance

Comparison of pseudo exception sample generation methods

The importance of decoupling learning And Number of reference images Compare