roc-ng/var

VAR

This is the official project website of our paper: Towards Video Anomaly Retrieval from Video Anomaly Detection: New Benchmarks and Model

Peng Wu, Jing Liu Senior Member, IEEE, Xiangteng He, Yuxin Peng Senior Member, IEEE, Peng Wang, and Yanning Zhang Senior Member, IEEE

UCF-Crime Captions

UCFCrime-AR Captions are released to facilitate future research.

Features

I3D features of UCFCrime-AR

I3D&VGGish features of XDViolence-AR

Highlight

• We introduce a new task named video anomaly retrieval to bridge the gap between the literature and real-world applications in terms of video anomaly analysis. To our knowledge, this is the first work moves towards VAR from VAD;
• We present two large-scale benchmarks, i.e., UCFCrime- AR and XDViolence-AR, based on public VAD datasets. The former is applied to video-text VAR, the latter is to video-audio VAR;
• We propose a model called ALAN, aiming at challenges in VAR, where anomaly-led sampling, video prompt based masked phrase modeling, and cross-modal align- ment are introduced for the attention of anomalous segments, enhancement of fine-grained associations, and multi-perspective match, respectively.

Citation

If you find this repo useful for your research, please consider citing our paper:

@article{wu2023towards,
  title={Towards Video Anomaly Retrieval from Video Anomaly Detection: New Benchmarks and Model},
  author={Wu, Peng and Liu, Jing and He, Xiangteng and Peng, Yuxin and Wang, Peng and Zhang, Yanning},
  journal={arXiv preprint arXiv:2307.12545},
  year={2023}
}

---

var's Issues

Model architecture release

Thank you for the excellent work you're doing. I'm currently researching VAR and was wondering when the full model architecture and weights will be released. Is there any possibility of getting early access?

I appreciate your time and consideration.