Our work tries to build robust anomaly detection models for surveillance videos while minimizing the need for human annotation by adopting sophisticated learning techniques (semi-supervision for weakly annotated data).
This work has been published in VISAPP 2022 and extended in my masters' thesis. You can find those resources at:
- VISAPP 2022 Published paper.
- Masters' Thesis.
- Thesis defense video (10th of September, 2022, Alexandria University): discussion in Arabic.
- Thesis defense slides: recommended as best place to start from.
This work uses the dataset from Sultani et al. 2018 and its baseline for benchmarks. Our work is also built on top of SlowFast to use its different vision models for feature extraction.
The following examples show how our model behaves in contrast to ground truth and baseline.
Please find installation instructions for the environment environment.sh. You may follow the instructions in download_dataset.s to download and prepare the datasets.
Please find the feature extraction, training, and inference instructions in the GETTING_STARTED.md document.
- Add libraries versions to the installation guide
- Add model weights
This work is released under the Apache 2.0 license.
If you find this project useful in your research, please use the following BibTeX entry for citation.
@inproceedings{el2022semi,
title={Semi-supervised Anomaly Detection for Weakly-annotated Videos.},
author={El-Tahan, Khaled and Torki, Marwan},
booktitle={VISIGRAPP (5: VISAPP)},
pages={871--878},
year={2022}
}