Multi-Scenario Anomaly Detection (MSAD) Dataset
Liyun Zhu1
Lei Wang1,2
Arjun Raj1
Tom Gedeon3
Chen Chen4
1Australian National University
2Data61/CSIRO
3Curtin University
4University of Central Florida




Overview: Our MSAD includes a diverse range of scenarios, both indoor and outdoor, featuring various objects, eg, pedestrians, cars, trains, etc. The first row shows different real-world common motions, while the second row demonstrates variations in weather and lighting conditions. The third row displays different moving objects. The last column shows human- and non-human-related anomalies.



Paper

MSAD Dataset

Code (Coming soon)


Dataset Abstract

We introduce a new dataset, Multi-Scenario Anomaly Detection (MSAD), comprising 14 distinct scenarios captured from various camera views. Our dataset has diverse motion patterns and challenging variations, such as different lighting and weather conditions, providing a robust foundation for training superior models. We conduct an in-depth analysis of recent representative models using MSAD and highlight its potential in addressing the challenges of detecting anomalies across diverse and evolving surveillance scenarios.


Dataset Overview

Total Videos

720

Domain

Multiple

#Human Related Anomaly Type

35

#Non-human Related Anomaly Type

20

#Scenarios

14

#View

∼500

Modality

RGB

Resolution

Multiple



Click to view extended details of our MSAD dataset. Our dataset has a broad range of both human-related and non-human-related anomalies, spanning multiple domains.
Main Anomaly Types Detailed Anomaly Types Domain
Human-related Assault Assault on street Crime
Assault in office Crime
Other assault Crime
Fighting Fighting on street Violence
Fighting in a restaurant Violence
Fighting in a shop Violence
Fighting in front of a door Violence
Fighting indoors Violence
Other fighting Violence
People Falling People falling to ground Pedestrian
People falling into pool Pedestrian
People falling from high places Pedestrian
People falling into subway Pedestrian
Other people falling Pedestrian
Robbery Shop robbery Crime
Office robbery Crime
Theft Crime
Car theft Crime
Other robbery Crime
Shooting Shooting on the road Crime
Shooting indoors Crime
Holding a gun Crime
Other shooting Crime
Traffic Accident Car falling Traffic
Car crash Traffic
Speeding Traffic
Car rushing into building Traffic
Car crash with people Traffic
Car crash with object Traffic
Car crash with train Traffic
Motorcycle crash Traffic
Other traffic accident Traffic
Vandalism Vandalizing glass Violence
Vandalizing door Violence
Other vandalism Violence
Non-human-related Explosion Street explosion Emergency
Firework explosion Emergency
Factory explosion Emergency
Indoor explosion Emergency
Other explosion Emergency
Fire Smoke Emergency
Factory fire Emergency
Building on fire Emergency
Bush fire Emergency
Other fire Emergency
Object Falling Strong wind Natural hazard
Object falling in home Emergency
Tree falling Emergency
Large objects falling Emergency
Glass falling Emergency
Other objects falling Emergency
Water Incident Flood Natural hazard
Water leakage Emergency
Heavy rain Natural hazard
Other water incidents Emergency

Click to view some graphical statistics of our MSAD dataset.



The statistics of MSAD dataset include: (left) a breakdown of main anomaly types and their respective percentages, (middle) a boxplot illustrating frame number variations across scenarios in MSAD training set, and (right) the distributions of train/test splits across scenarios.


Dataset Evaluation

Based on the experimental results, we can deduce that a model trained on intricate real-world scenarios exhibits superior generalization. This stems from the fact that real-world models are frequently influenced by the surrounding environ ment, encompassing elements like fluctuating traffic patterns, dynamic electronic displays, and the movement of trees in the wind. The model must discern the nuances of anomaly detection within a dynamic environment and comprehend the dynamics of objects and/or performing subjects within it. MSAD dataset provides a comprehensive representation of real-world scenarios.

Generalizability and adaptability

Test view Train FSAD Train FSAD SA2D (ours)
Micro Macro Micro Macro Micro Macro
ShT-v1 ShT (7 views) 61.36 55.34 MSAD 63.74 62.92 68.96 77.89
ShT-v3 ShT (7 views) 26.51 26.58 MSAD 64.39 62.56 67.59 73.43
ShT-v5 ShT (7 views) 53.40 53.32 MSAD 55.04 54.63 55.74 54.02
ShT-v6 ShT (7 views) 78.36 78.27 MSAD 70.26 71.02 75.47 72.35
ShT-v8 ShT (7 views) 50.02 52.54 MSAD 59.97 57.45 60.85 61.52
Experimental results on single-scenario evaluations. On ShanghaiTech (ShT), only 7 views are used during training and the rest views are individually used for testing. The notation ShT-v∗ denotes the use of different camera views.

Cross-scenario evaluation

Train Test Micro Macro
ShT UCSD Ped2 57.38 58.36
ShT CUHK Avenue 69.98 78.32
ShT MSAD 63.92 64.92
MSAD UCSD Ped2 70.35 65.74
MSAD CUHK Avenue 79.57 84.49
MSAD MSAD 69.96 69.60
Evaluations on cross-scenario setups. We use FSAD and SA2D (ours) for training on ShanghaiTech (ShT) and MSAD, respectively.


Download

Please read the terms and conditions here and complete the below linked request form. Upon review, you will be notified the outcome of your request via the provided email.

v1.0



MSAD Dataset


Paper


Advancing Video Anomaly Detection: A Concise Review and a New Dataset


Cite

@misc{zhu2024advancingvideoanomalydetection,
      title={Advancing Video Anomaly Detection: A Concise Review and a New Dataset}, 
      author={Liyun Zhu and Lei Wang and Arjun Raj and Tom Gedeon and Chen Chen},
      year={2024},
      eprint={2402.04857},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2402.04857}, 
}


Contact

Please contact the following people for any enquires related to the dataset or the paper.