Advancing Anomaly Detection: An Adaptation Model and a New Dataset
Paper • 2402.04857 • Published
MSAD Dataset Access Request Form
Please read the terms and conditions below carefully and complete all fields accurately. Access is available only for non-commercial academic research and will be granted only after review and approval by the dataset maintainers. Submit your request using this form. For questions, email Liyun.Zhu@alumni.anu.edu.au and CC time.lab@griffith.edu.au and l.wang4@griffith.edu.au. If you have no additional comments, enter N/A.
MSAD Dataset Terms and Conditions
Permitted Use: The Dataset is to be used exclusively for academic and research purposes. Commercial use, reproduction, distribution, or sale of the Dataset or any derivative works is strictly prohibited.
Confidentiality: The Dataset must not be disclosed, shared, or disseminated to any third party without prior written consent from the Provider. All data must be handled with appropriate confidentiality and security measures.
Attribution: Any publications, presentations, or reports that result from the use of the Dataset must acknowledge the Provider and include a citation to the original dataset paper.
Prohibited Activities: You shall not attempt to re-identify any individuals or entities within the Dataset. You shall not use the Dataset to train or develop any models intended for surveillance or monitoring outside of research contexts.
Data Security: You must implement appropriate technical and organizational measures to protect the Dataset against unauthorized access, loss, alteration, or disclosure.
Compliance with Laws: You agree to comply with all applicable laws and regulations related to data protection and privacy in handling the Dataset.
Termination: The Provider reserves the right to terminate this agreement and require you to delete all copies of the Dataset if any terms of this agreement are violated.
Log in or Sign Up to review the conditions and access this dataset content.
MSAD: Multi-Scenario Anomaly Detection Dataset
A video anomaly detection benchmark introduced at NeurIPS 2024, Datasets and Benchmarks Track.
Overview
Multi-Scenario Anomaly Detection (MSAD) is a real-world RGB video benchmark for studying anomaly detection across diverse indoor and outdoor environments. It was introduced in “Advancing Video Anomaly Detection: A Concise Review and a New Dataset” by Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, and Chen Chen.
MSAD covers 14 scenarios, multiple camera viewpoints, varied motion patterns, and changes in lighting and weather. Its human-related and non-human-related anomalies support research into how models generalize across environments. See the paper for the research motivation and analysis.
Dataset Composition
The original benchmark contains 720 videos: 480 normal videos and 240 abnormal videos. The video packages prepared for this release are:
| Directory | MP4 files | Contents |
|---|---|---|
MSAD_normal_training/ |
360 | Normal training videos, grouped by scenario. |
MSAD_normal_testing_videos/ |
120 | Normal testing videos. |
MSAD_anomaly/ |
240 | Abnormal videos, grouped by anomaly category. |
VALU_MSAD_subset_2026/MSAD_split_mask_test_videos/ |
361 | Processed versions of the original normal testing and abnormal videos, prepared for VALU. |
The original MSAD videos are the first three packages. The VALU package overlaps with the original test data and does not add independent samples to the 720-video benchmark.
Scenarios and Anomaly Categories
The 14 scenarios are Front Door, Highway, Mall, Office, Park, Parking Lot, Pedestrian Street, Restaurant, Road, Shop, Sidewalk, Street Highview, Train, and Warehouse.
The 11 main anomaly categories are Assault, Explosion, Fighting, Fire, Object Falling, People Falling, Robbery, Shooting, Traffic Accident, Vandalism, and Water Incident. The project website describes their finer-grained subtypes and provides examples and statistics.
Video Package Layout
MSAD/
├── README.md
├── MSAD_normal_training/
│ ├── frontdoor/
│ ├── highway/
│ └── ...
├── MSAD_normal_testing_videos/
│ └── *.mp4
├── MSAD_anomaly/
│ ├── Assault/
│ ├── Explosion/
│ └── ...
└── VALU_MSAD_subset_2026/
├── README.md
└── MSAD_split_mask_test_videos/
└── *.mp4
Processed MSAD Subset for VALU
The VALU package represents the 360 source videos used for testing in Protocol i: 120 normal videos and 240 abnormal videos. It includes watermark processing where applicable and clip segmentation for some videos with scene or camera changes. Normal training videos are not included in this processed package.
Use the original video packages when reproducing MSAD benchmarks. For experiments using the processed package, consult its README and the VALU repository for preprocessing and annotation guidance, and acknowledge both MSAD and VALU.
Evaluation Protocols
MSAD defines two evaluation protocols:
| Protocol | Training videos | Testing videos | Setting |
|---|---|---|---|
| i | 360 normal | 120 normal + 240 abnormal | Self-supervised anomaly detection. |
| ii | 360 normal + 120 abnormal | 120 normal + 120 abnormal | Weakly supervised anomaly detection, with video-level training labels. |
For Protocol ii, use the published split lists to select the abnormal training and testing videos. The MSAD_anomaly/ folder contains all 240 abnormal videos; its category folders do not define this split. Protocol details are available on the project website.
The benchmark repository implements Protocol ii and provides RTFM and MGFN examples. Its dataset preparation section links to extracted I3D and Video-Swin Transformer features, training/testing lists, and pretrained checkpoints. Follow those lists and the accompanying evaluation setup when comparing results.
Access and Download
Raw video access requires review and approval. Complete the Hugging Face access request form accurately and review the existing terms and conditions before requesting or using the data. The academic and research use requirements are also stated in the source repository.
After access has been approved, install the Hugging Face CLI and log in with your approved account:
python3 -m pip install --upgrade huggingface_hub
hf auth login
Download the repository:
hf download Liyun1009/MSAD --repo-type dataset --local-dir ./MSAD
Or download an individual package, for example the normal testing videos:
hf download Liyun1009/MSAD \
--repo-type dataset \
--include "MSAD_normal_testing_videos/**" \
--local-dir ./MSAD
The Hugging Face CLI documentation explains installation, authentication, and download options.
Citation
Please cite the dataset paper when using MSAD:
@inproceedings{msad2024,
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},
booktitle = {The Thirty-eighth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
year = {2024}
}
Contact
For dataset access questions, contact Liyun.Zhu@alumni.anu.edu.au, with time.lab@griffith.edu.au and l.wang4@griffith.edu.au in CC.
license: cc-by-4.0
- Downloads last month
- 33