Consistency-based Self-Supervised Learning for Temporal Anomaly Localization

This repository contains Pytorch code for the WCPA ECCV22 paper "Consistency-based Self-Supervised Learning for Temporal Anomaly Localization" [arXiv]

@inproceedings{panariello2022consistency,
    title = {Consistency-based Self-supervised Learning for Temporal Anomaly Localization},
    author = {Panariello, Aniello and Porrello, Angelo and Calderara, Simone and Cucchiara, Rita},
    booktitle = {Proceedings of the European Conference on Computer Vision (ECCV) Workshops},
    month = {October},
    year = {2022},
}

Installation Note

Tested with Python 3.8.13 on Ubuntu (22.04).

• Setup an empty pip environment
• Install packages using pip install -r requirements.txt
• Place dataset in ./data/ Download Link
• Run main.py

Please note that if you're running the code from Pycharm (or another IDE) you may need to manually set the working path to PROJECT_PATH

Run python main.py to train the model.

Improvements over the original paper

• Removed smoothness loss as it was in conflict with the alignment loss. This leads to better and more stable results.
• Add support for gated attention [1] leading to a +3% improvement in AP frame-level.

To replicate the results of the paper, run:

python main.py --batch-size 8 --alpha 2e-8 --gamma 0.5 --no-gated-attention

---

References

[1] Ilse, Maximilian and Tomczak, Jakub and Welling, Max. Attention-based deep multiple instance learning. International conference on machine learning. PMLR, 2018.