Reproduce “Asymmetric Student-Teacher Networks for Industrial Anomaly Detection” on OpenResearch

About this paper

Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, Bastian Wandt

A new Asymmetric Student-Teacher (AST) network addresses the undesired generalization issue in student-teacher anomaly detection by employing a bijective Normalizing Flow as a teacher and a conventional network as a student. This approach achieved a 99.2% mean detection AUROC on MVTec AD and a 93.7% mean AUROC on MVTec 3D-AD, setting new state-of-the-art results on these industrial datasets.

Key insights

• The architectural similarity in existing student-teacher networks causes undesired generalization, where the student can mimic the teacher's output even for out-of-distribution anomalous inputs.
• Employing an asymmetric architecture with a bijective Normalizing Flow as the teacher and a conventional network as the student prevents the student from generalizing to anomalies, leading to robust anomaly detection.
• Using the Normalizing Flow as a teacher within a student-teacher setup provides a more reliable anomaly score, compensating for the inherent ambiguities and potential misestimations of pure likelihood-based anomaly scores from NFs.

Summary by alphaXiv.