Lightning Fast Video Anomaly Detection via Multi-Scale Adversarial Distillation

July 10, 2024 ยท View on GitHub

Computer Vision and Image Understanding (Official Repository)

Overview

Fast-AED is a high-performance frame-level anomaly detection model for videos. It offers exceptional speed while maintaining competitive accuracy, making it ideal for real-time applications. The model achieves this by distilling knowledge from multiple precise object-level teacher models and employing a unique combination of standard and adversarial distillation techniques.

For more details please read our full paper.

Features

  • Knowledge Distillation: Transfer of information from accurate object-level teacher models to a fast student model.

  • Adversarial Distillation: Joint application of standard and adversarial distillation to enhance the fidelity of anomaly maps.

  • Adversarial Discrimination: Introduction of adversarial discriminators for each teacher model to distinguish between target and generated anomaly maps.

  • Benchmark Performance: Extensive experiments on Avenue, ShanghaiTech, and UCSD Ped2 datasets showcase the model's superiority, outperforming competitors by over 7 times in speed and 28 to 62 times faster than object-centric models.

  • Unprecedented Speed: Operating at an outstanding 1480 frames per second (FPS), Fast-AED achieves the best trade-off between speed and accuracy.

License

The source code and models are released under the Creative Common Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license.

Citation

Please cite our work if you use any material released in this repository.

@article{Croitoru-CVIU-2024,
  author    = {Croitoru, Florinel-Alin and Ristea, Nicolae-Catalin and Dascalescu, Dana and Ionescu, Radu Tudor and Khan, Fahad Shahbaz and Shah, Mubarak},
  title     = "{Lightning Fast Video Anomaly Detection via Multi-Scale Adversarial Distillation}",
  journal   = {Computer Vision and Image Understanding},
  year      = {2024},
  }

Getting Started

Prerequisites

  • Python 3.6+
  • PyTorch 1.2+
  • CUDA Toolkit (for GPU support)

Installation

Clone the repository:

git clone https://github.com/ristea/fast-aed.git
cd fast-aed