Env-Aware Anomaly Detection: Ignore Style Changes, Stay True to Content!

TL;DR

This paper formalizes and benchmarks environment-aware anomaly detection in visual data, demonstrating improved performance over traditional methods by incorporating environment labels and generating positive samples accordingly.

Contribution

It introduces a novel formalization and benchmark for environment-aware anomaly detection in visual data, and proposes an extension for contrastive learning that leverages environment labels.

Findings

Environment-aware methods outperform ERM in distribution-shift scenarios.

The proposed positive sample generation improves ERM baseline by 8.7%.

First approach for environment-aware anomaly detection in visual datasets.

Abstract

We introduce a formalization and benchmark for the unsupervised anomaly detection task in the distribution-shift scenario. Our work builds upon the iWildCam dataset, and, to the best of our knowledge, we are the first to propose such an approach for visual data. We empirically validate that environment-aware methods perform better in such cases when compared with the basic Empirical Risk Minimization (ERM). We next propose an extension for generating positive samples for contrastive methods that considers the environment labels when training, improving the ERM baseline score by 8.7%.