Official implementation of "LUSS-AE"

This is the implementation of our paper "Leveraging Unsupervised and Self-Supervised Learning for Video Anomaly Detection" (VISAPP 2023).

Cloning repository

git clone https://github.com/devashishlohani/luss-ae_vad

Dependencies

• Python 3.6
• PyTorch >= 1.7.0
• Numpy
• Sklearn

Use conda to install dependencies from requirements.yml

Datasets

• USCD Ped2 [dataset]
• CUHK Avenue [dataset]
• ShanghaiTech [dataset]

Download the datasets and put into dataset folder, like ./dataset/ped2/, ./dataset/avenue/, ./dataset/shanghai/

Training

python train.py --dataset_type shanghai

Select --dataset_type from ped2, avenue, or shanghai.

For more details, check train.py

Testing

• First fetch the pre-trained models from drive
• Place the model in exp folder like ./exp/shanghai/pre_trained/model_best.pth
• Run the test script as shown below

python test.py --test_batch_size 4 --num_workers_test 4 --dataset_type shanghai --model_path exp/shanghai/pre_trained/model_best.pth 

For more details, check arguments in test.py

Bibtex

@inproceedings{lohani2023leveraging,
  title={Leveraging Unsupervised and Self-Supervised Learning for Video Anomaly Detection},
  author={Lohani, Devashish and Crispim-Junior, Carlos F and Barth{\'e}lemy, Quentin and Bertrand, Sarah and Robinault, Lionel and Tougne, Laure},
  booktitle={18th International Conference on Computer Vision Theory and Applications},
  volume={5},
  pages={132--143},
  year={2023},
  organization={SCITEPRESS-Science and Technology Publications}
}