Kart Laura ; Cohen Niv

Abstract

In recent years, many works have addressed the problem of findingnever-seen-before anomalies in videos. Yet, most work has been focused ondetecting anomalous frames in surveillance videos taken from security cameras.Meanwhile, the task of anomaly detection (AD) in videos exhibiting anomalousmechanical behavior, has been mostly overlooked. Anomaly detection in suchvideos is both of academic and practical interest, as they may enable automaticdetection of malfunctions in many manufacturing, maintenance, and real-lifesettings. To assess the potential of the different approaches to detect suchanomalies, we evaluate two simple baseline approaches: (i) Temporal-pooledimage AD techniques. (ii) Density estimation of videos represented withfeatures pretrained for video-classification. Development of such methods calls for new benchmarks to allow evaluation ofdifferent possible approaches. We introduce the Physical Anomalous Trajectoryor Motion (PHANTOM) dataset, which contains six different video classes. Eachclass consists of normal and anomalous videos. The classes differ in thepresented phenomena, the normal class variability, and the kind of anomalies inthe videos. We also suggest an even harder benchmark where anomalous activitiesshould be spotted on highly variable scenes.