A Discriminative Framework for Anomaly Detection in Large Videos

摘要: We address an anomaly detection setting in which training sequences are unavailable and anomalies scored independently of temporal ordering. Current algorithms based on the classical density estimation approach learning high-dimensional models finding low-probability events. These sensitive to order appear require either data or early context assumptions that do not hold for longer, more complex videos. By defining as examples can be distinguished from other same video, our definition inspires a shift approaches simple discriminative learning. Our contributions include novel framework is (1) independent ordering anomalies, (2) unsupervised, requiring no separate sequences. show algorithm achieve state-of-the-art results even when we adjust by removing standard datasets.