Continual-MEGA: A large-scale benchmark for generalizable continual anomaly detection

Anomaly detection is essential for visual inspection systems that require reliable identification of defective or abnormal samples. In real-world deployment, however, object categories and defect types can change over time, making static evaluation settings insufficient for assessing long-term adaptability and generalization. To address this limitation, we introduce a new benchmark for continual learning in anomaly detection. Our benchmark, Continual-MEGA, expands existing evaluation settings by integrating diverse public datasets with our newly proposed dataset, ContinualAD. Beyond standard continual learning settings, we additionally propose a scenario that evaluates zero-shot generalization to unseen classes not encountered during continual adaptation. This reflects recent advances in continual zero-shot research and highlights its practical significance. This setting introduces a new agenda for the anomaly detection field, and we conduct extensive evaluations of various existing anomaly detection algorithms designed for continual or zero-shot scenarios, as well as our proposed baseline methods. From our experiments, we derive three key findings: (1) existing methods exhibit significant limitations, particularly in pixel-level defect localization, (2) the proposed ContinualAD dataset is effective for the proposed benchmarking scenario, and (3) our baseline method suggests a promising direction for designing CLIP-based continual and generalizable frameworks through simple adaptation combined with feature synthesis.