Rethinking benchmarking framework of self-supervised learning approaches for anomaly localization

Localizing defects in products is a critical component of industrial pipelines in manufacturing, retail, and many other industries to ensure consistent delivery of
the highest quality products. Automated anomaly localization systems leveraging
computer vision have the potential to replace laborious and subjective manual
inspection of products. Recently, there have been tremendous efforts in the domain
of anomaly localization investigating self-supervised learning methods. However,
despite the advancements, there is still a gap between research and deployment of
those methods to real-world production environment. It is important to develop
an industry-friendly benchmarking framework to understand the performance of
models in a generalizable product-agnostic manner. We present a new anomaly
localization benchmarking framework that maps a product/defect type combination
to higher level descriptive abstractions capturing similar characteristics. We propose
efficient training and inference schemes considering different aspects, including an
ablation study of threshold estimation techniques. To the best of our knowledge,
this is the first anomaly localization work on developing a benchmarking framework
focusing on real-world use.