PHANTOM (Physical Anomalous Trajectory or Motion (PHANTOM))

近年来，许多作品都解决了在视频中发现从未见过的异常的问题。然而，大多数工作都集中在检测从安全摄像机拍摄的监控视频中的异常帧。同时，表现出异常机械行为的视频中的异常检测 (AD) 任务大多被忽视了。此类视频中的异常检测具有学术和实际意义，因为它们可以自动检测许多制造、维护和现实生活中的故障。为了评估检测此类异常的不同方法的潜力，我们评估了两种简单的基线方法：（i）时间池图像 AD 技术。 (ii) 用视频分类预训练的特征表示的视频的密度估计。开发此类方法需要新的基准，以评估不同的可能方法。我们介绍了物理异常轨迹或运动 (PHANTOM) 数据集，其中包含六个不同的视频类。每个类由正常和异常视频组成。这些类别在呈现的现象、正常类别的可变性和视频中的异常类型方面有所不同。我们还提出了一个更难的基准，应该在高度可变的场景中发现异常活动。

In recent years, numerous works have addressed the problem of detecting unseen anomalies in videos. However, most of these works focus on detecting anomalous frames in surveillance footage captured by security cameras. Meanwhile, the anomaly detection (AD) task in videos exhibiting abnormal mechanical behaviors has largely been overlooked. Anomaly detection in such videos holds both academic and practical significance, as they can automatically detect numerous faults in manufacturing, maintenance, and real-world scenarios. To evaluate the potential of different methods for detecting such anomalies, we assess two simple baseline methods: (i) temporal-pooled image AD techniques; (ii) density estimation on video features pre-trained for video classification. Developing such methods requires new benchmarks to evaluate diverse potential approaches. We introduce the Physical Anomaly Trajectory or Motion (PHANTOM) dataset, which encompasses six distinct video classes. Each class consists of both normal and anomalous videos. These classes differ in terms of the presented phenomena, the variability of their normal samples, and the types of anomalies present in the videos. We also propose a more challenging benchmark that requires detecting anomalous activities in highly variable scenarios.

搜集汇总

背景与挑战

背景概述

PHANTOM数据集是一个用于检测视频中物理异常轨迹或运动的基准，包含六个不同类别的正常和异常视频，旨在评估机械行为异常检测方法。该数据集由耶路撒冷希伯来大学于2021年发布，以弥补现有研究在监控视频异常检测之外的不足。