Revisiting Anomaly Localization Metrics

David Zimmerer, Klaus Maier-Hein

Keywords: Anomaly Localization, Anomaly Detection, Metrics

Abstract:

An assumption-free, disease-agnostic pathology detector and segmentor could often be seen as one of the holy grails of medical image analysis.
Building on this promise, un-/weakly supervised anomaly localization approaches, which aim to model normal/healthy samples using data and then detect anything deviant from this (i.e., anything abnormal), have gained popularity.
However, being an upcoming area in between image segmentation and out-of-distribution detection, most approaches have adapted their evaluation setup and metrics from either field and thus might have missed peculiarities inherent to the anomaly localization task.
Here, we revisit the anomaly localization setup, discuss and analyse the properties of the often used metrics, show alternative metrics inspired from instance segmentation and compare the metrics across multiple setting and algorithms.
Overall, we argue that the choice of the metric is use-case dependent, however, the Soft Instance IoU shows significant promise going forward.

Submission Number: 82