Anomaly detection is a growing field where the goal is to distinguish between normal and abnormal samples. Supervised methods are ineffective as there is a lack of sufficient representative samples of the anomalous class. To address this problem, the anomaly detection algorithms rely solely on the normal images during training. This makes it well-suited for applications in industry, medicine, and security where a clearly defined anomalous class is often lacking. With the increasing number of publications, Anomalib provides a unified library for benchmarking algorithms and toolings for research. Since it was made public, Anomalib has been cited by papers published in a lot of venues CVPR and had a booth of its own in CVPR 2023.
In this talk, we explain how Anomalib is designed in terms of modularity to allow adding any new algorithm, the tooling available for reproducibility and benchmarking, and finally, optimizing the models for deployment on edge devices.
Ashwin Vaidya graduated from the University of Groningen in August 2021 with a Master’s in Artificial Intelligence. Since then I have been working full-time with the Anomaly Detection team at Intel. While my primary research interest is in the field of Reinforcement Learning, my current interests also include one-class, few-shot, and continual learning.
Dick Ameln is one of the founding members and main contributors of Anomalib. He graduated from University of Groningen in 2020 with a double master degree in Human Movement Sciences and Artificial Intelligence. He is currently active as AI research engineer at Intel, where his main focus is the design and implementation of anomaly detection algorithms.