Weakly Supervised Video Anomaly Detection with Temporal and Abnormal Information

Abstract

AbstractWeakly supervised video anomaly detection is to distinguish anomalies from normal scenes and events in videos, under the setting that we only know whether there are abnormal events in a video, but the specific occurrence time of abnormal events is unknown. It is generally modeled as a MIL (multiple instance learning) problem, where video-level labels are provided to train an anomaly detector to obtain frame-level labels for videos. However, most existing methods generally overlook temporal information in abnormal videos (positive bags), and only use one sample (snippet) in the positive bag to train. The positive bag may include more useful information with high possibility. Therefore, we propose the Weakly Supervised Video Anomaly Detection Approach with Temporal and Positive Features, which consider both the temporal information and more positive samples for video anomaly detection. Its contributions can be summarized as follows: (1) we consider more temporal information and introduced the attention mechanism in our network, we use both local and global snippets’ features to enhance the temporal representation ability of these features. (2) We use more positive (abnormal) samples and its features in bags to train our model, so that more complementary and relevant information will make our model more robust and effective. (3) We consider not only the differences between normal samples and abnormal samples but also between abnormal samples and abnormal samples, which can help our proposed approach to excavate positive (abnormal) samples’ information more efficiently and adequately. Experimental results demonstrate the effectiveness of our proposed methods in the UCF-Crime and ShanghaiTech dataset.KeywordsVideo anomaly detectionWeakly supervisionMultiple instance learningTemporal features