Training

1. a projection layer directly after each intermediate teacher block to provide the student network with a compact, abnormal-free representation.
   1. L_{SSOT}: Self-supervised optimal transport loss for compact
      1. projected normal features should be close to each other.
   2. L_{recon}: Reconstruction Loss for abnormal-free
      1. projected anomaly features should be like projected normal features.
   3. L_{Con}: Contrast Loss
      1. projected normal features should be dislike anomaly features.
2. one-class embedding (OCBE) module
3. L_{KD}: loss for knowledge transfer;

Inferences

1. Revisiting Reverse Distillation for Anomaly Detection, 23, ranked 16
   1. Ranked #2 on Anomaly Detection on MVTEC AD textures