# Approaches Toward Physical and General Video Anomaly Detection

**Authors:** Laura Kart, Niv Cohen  
**arXiv:** [2112.07661](https://arxiv.org/abs/2112.07661)  
**Date:** 2021-12-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2112.07661v1/

## Key points

- Detecting never-seen-before anomalies in videos is a focus of research.
- Anomaly detection (AD) in videos exhibiting anomalous mechanical behavior has been largely ignored.
- Automatic detection of malfunctions in manufacturing, maintenance, and real-life settings can be enabled through AD in such videos.
- Two simple baseline techniques were assessed to detect anomalies in these types of videos: temporal-pooled image AD techniques and density estimation of videos represented with features pretrained for video-classification.
- The authors introduce the Physical Anomalous Trajectory or Motion (PHANTOM) dataset that contains six distinct video classes to enable evaluation of various possible approaches.
- The PHANTOM dataset provides a valuable resource for evaluating different approaches to achieve this goal which can help improve automatic detection of malfunctions in manufacturing processes and other real-life settings.

## Summary

The problem of detecting never-seen-before anomalies in videos has been a focus of research in recent years. However, most studies have concentrated on identifying anomalous frames in surveillance videos captured by security cameras. The task of anomaly detection (AD) in videos exhibiting anomalous mechanical behavior has been largely ignored, despite its academic and practical significance. Automatic detection of malfunctions in manufacturing, maintenance, and real-life settings can be enabled through AD in such videos. To evaluate the potential of different approaches to detect anomalies in these types of videos, two simple baseline techniques were assessed: (i) Temporal-pooled image AD techniques and (ii) Density estimation of videos represented with features pretrained for video-classification. Development of such methods necessitates new benchmarks to enable evaluation of various possible approaches. To address this need, the authors introduce the Physical Anomalous Trajectory or Motion (PHANTOM) dataset that contains six distinct video classes. Each class comprises normal and anomalous videos that differ based on presented phenomena, normal class variability, and the kind of anomalies present in the videos. Additionally, an even more challenging benchmark is suggested where anomalous activities must be detected on highly variable scenes. Overall, this study highlights the importance of developing effective AD techniques for detecting anomalies in videos exhibiting anomalous mechanical behavior. The PHANTOM dataset provides a valuable resource for evaluating different approaches to achieve this goal which can help improve automatic detection of malfunctions in manufacturing processes and other real-life settings.