Reproduce “Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xincheng Yao, Ruoqi Li, Jing Zhang, Jun Sun, Chongyang Zhang

Researchers from Shanghai Jiao Tong University developed Boundary Guided Anomaly Detection (BGAD), a method for supervised anomaly detection that uses limited known anomalies to enhance discriminability without bias towards seen anomalies. It achieved 99.3% image-level AUROC on MVTecAD and superior generalization to unseen anomalies across benchmarks.

Key insights

  • Defining explicit decision boundaries based purely on the normal data distribution effectively prevents models from overfitting to the specific characteristics of limited known anomalies.
  • The semi-push-pull contrastive loss, which pushes abnormal features only to a specific margin beyond the normal boundary, helps maintain generalization by avoiding over-optimization on seen anomalies.
  • Integrating density estimation of normal data with boundary-guided contrastive learning creates a feature space that is both compact for normal instances and discriminative for diverse anomalies.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.