Env-Aware Anomaly Detection: Ignore Style Changes, Stay True to Content!
This work addresses anomaly detection in visual data with distribution shifts, which is an incremental advancement for computer vision applications.
The paper tackles unsupervised anomaly detection under distribution shift by formalizing a benchmark on the iWildCam dataset, showing that environment-aware methods outperform basic Empirical Risk Minimization (ERM) and proposing an extension that improves the ERM baseline by 8.7%.
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%.