Consistency Bidirectional GAN (CBiGAN)

CBiGAN: a combined model that generalizes Bidirectional GANs (BiGANs) and AutoEncoders, applied to anomaly detection in images. The repo provides training and evaluation code for the MVTecAD anomaly detection benchmark.

Also provides a TensorFlow2 implementation of BiGAN following the Wasserstein GAN (WGAN) formulation.

Getting started

You need:

• Python 3
• Tensorflow 2.4.0
• packages in requirements.txt

You can use the Dockerfile to build an image.

Train on MVTec-AD

Download the whole MVTec-AD dataset and extract into data/mvtec-ad.

Check out the train.py script for training parameters:

python train.py -h

Reference

Combining GANs and AutoEncoders for Efficient Anomaly Detection. Fabio Carrara, Giuseppe Amato, Luca Brombin, Fabrizio Falchi, Claudio Gennaro. In 2020 25th International Conference on Pattern Recognition (ICPR) (pp. 3939-3946). IEEE. [arXiv, DOI]

@inproceedings{carrara2021combining,
  title={Combining gans and autoencoders for efficient anomaly detection},
  author={Carrara, Fabio and Amato, Giuseppe and Brombin, Luca and Falchi, Fabrizio and Gennaro, Claudio},
  booktitle={2020 25th International Conference on Pattern Recognition (ICPR)},
  pages={3939--3946},
  year={2021},
  organization={IEEE}
}