EventVAD: Training-Free Event-Aware Video Anomaly Detection (Accepted by ACM Multimedia 2025)

If our project helps you, please give us a star ⭐ on GitHub to support us. 🙏🙏

News

• [2025.07.09] 🔥 We initiated this project and made both the code and data publicly available as open source.
• [2025.07.06] Our paper was accepted by ACM Multimedia 2025.

Datasets

This project uses the following datasets for evaluation:

• UCF-Crime: A large-scale dataset for real-world surveillance video anomaly detection, containing 1900 long untrimmed videos with 13 real-world anomalies.
  Download link

• XD-Violence: A large-scale audio-visual dataset for violence detection in videos, comprising 4754 untrimmed videos covering 6 categories of anomalies.
  Download link

Place the datasets in the directories specified by the code (e.g., src/event_seg/videos for input videos).

Installation

This project requires Python 3.10. It is recommended to use a virtual environment to manage dependencies.

1. Clone the repository:

   git clone https://github.com/YihuaJerry/EventVAD.git

2. Install the packages required for event splitting:

   conda create -n event_seg python=3.10
   conda activate event_seg
   cd src/event_seg
   pip install -r requirements.txt

3. Install the packages required for video scoring:

   conda create -n score python=3.10
   conda activate score
   cd src/score
   pip install -r requirements.txt

Note: Make sure to activate the virtual environment before installing dependencies. The 'evaluate.py' script relies on the library in 'score'.

Usage

To run the entire pipeline, follow these steps:

1. Event Segmentation:

   • Navigate to the event segmentation directory:

     cd src/event_seg

   • Run the main script:

     conda activate event_seg
     python main.py

2. Scoring:

   • Navigate to the scoring directory:

     cd src/score

   • Run the scoring script:

     conda activate score
     python event_score.py

3. Evaluation:

   • Run the evaluation script:

     cd src
     conda activate score
     python evaluate.py

Note: Ensure that the UCF-Crime and XD-Violence datasets are prepared and placed in the appropriate directories (e.g., src/event_seg/videos for input videos).

Main Results

UCF-Crime

XD-Violence

📑 Citation

If you use this code or the EventVAD method in your research, please cite the following paper:

@article{shao2025eventvad,
  title={Eventvad: Training-free event-aware video anomaly detection},
  author={Shao, Yihua and He, Haojin and Li, Sijie and Chen, Siyu and Long, Xinwei and Zeng, Fanhu and Fan, Yuxuan and Zhang, Muyang and Yan, Ziyang and Ma, Ao and others},
  journal={arXiv preprint arXiv:2504.13092},
  year={2025}
}