Reproduce “BAAF: Universal Transformation of One-Class Classifiers for Unsupervised Image Anomaly Detection” on OpenResearch

About this paper

Declan McIntosh, Alexandra Branzan Albu

Bootstrap Aggregation Anomaly Filtering (BAAF) is proposed as a regularization technique to transform existing One-Class Classification (OCC) anomaly detectors into fully unsupervised methods. This enables state-of-the-art performance on datasets with corrupted training data, including the first unsupervised method for logical anomaly detection, by robustly filtering anomalous samples without modifying the OCC model architecture.

Key insights

• The inherent rarity and heterogeneity of anomalies can be leveraged to effectively filter anomalous samples from noisy training data through collaborative predictions.
• A universal transformation is possible for diverse OCC anomaly detection methods, enabling them to operate in fully unsupervised settings by iteratively purifying their training data.
• Advancements in OCC architectures can directly translate into improvements in unsupervised anomaly detection when combined with a robust filtering mechanism like BAAF.

Summary by alphaXiv.