sy00n / DL_paper_review

issues에 논문 요약
0 stars 0 forks source link

[33] Prototypical Residual Networks for Anomaly Detection and Localization #38

Open sy00n opened 11 months ago

sy00n commented 11 months ago

이 논문 핵심

image

Multi-scale Prototypes

3.3 Multi-size Self-Attention

Anomaly Generation Strategies

  1. Extened anamalies(EA) : normal image에다가 seen anomaly로부터 augmented anomalous regions를 위치시키는 방식 (in-distribution anomalies)
  2. Simulated anomalies(SA) : seen anomalies에 대한 지식 없이 normal sample로부터 만들기

Extended Anomalies

Training and Inference

decoder가 anomaly sore map을 만들면 focal loss랑 sooth L1 loss로 최적화함.