UMAD: University of Macau Anomaly Detection Benchmark Dataset

In this paper, we propose UMAD, the first benchmark dataset for the anomaly detection with reference in the context of patrol robot scenes. To address the challenge of aligning the reference and query pairs, we propose an adaptive warping method, which allows for accurate comparison between the image pairs. Unlike anomaly detection without reference approaches that face difficulties in handling the diverse range of anomalies, anomaly detection with reference offers an alternative by leveraging the change detection between the reference and query image pairs. By utilizing the UMAD dataset, we have conducted evaluations of several state-of-the-art scene change detection methods to identify anomalies and establish a baseline performance. We intend that UMAD will facilitate the development of more effective anomaly detection with reference methods, thereby bridging the existing gap in this research domain.