Reproduce “UBnormal: New Benchmark for Supervised Open-Set Video Anomaly Detection” on OpenResearch

About this paper

Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, Mubarak Shah

The paper introduces UBnormal, a novel benchmark dataset and "supervised open-set" formulation for video anomaly detection. It provides pixel-level abnormal annotations for training while maintaining disjoint anomaly types for testing, enabling models to improve real-world VAD performance, achieving up to 93.0% micro-averaged AUC on Avenue.

Key insights

• Providing pixel-level abnormal training data, even in a synthetic and open-set context, consistently benefits state-of-the-art anomaly detection models.
• UBnormal's challenging nature, evidenced by modest baseline performance (e.g., 68.5% micro-averaged AUC), makes it a valuable resource for developing more robust and generalizable VAD systems.
• Integrating UBnormal data, especially when coupled with domain adaptation techniques like CycleGAN, enhances the performance of state-of-the-art frameworks on established real-world VAD datasets.

Summary by alphaXiv.