Localizing Anomalies From Weakly-Labeled Videos

Video anomaly detection under video-level labels is currently a challenging task. Previous works have made progresses on discriminating whether video sequencecontains anomalies. However, most of them fail to accurately localize the anomalous events within videos in temporal domain. In this paper, we propose Weakly Supervised Anomaly Localization (WSAL) method focusing temporally localizing segments videos. Inspired by appearance difference videos, evolution adjacent evaluated for localization segments. To end, high-order context encoding model proposed not only extract semantic representations but also measure dynamic variations so that could be effectively utilized. addition, order fully utilize spatial information, immediate semantics are directly derived from segment representations. The as well semantics, efficiently aggregated obtain final scores. An enhancement strategy further deal with noise interference and absence guidance detection. Moreover, facilitate diversity requirement benchmarks, collect new traffic (TAD) dataset which specifies conditions, differing greatly current popular evaluation benchmarks.Extensive experiments conducted verify effectiveness different components, our achieves state-of-the-art performance UCF-Crime TAD datasets.