[AAAI2025] Filter or Compensate: Towards Invariant Representation from Distribution Shift for Anomaly Detection

Link to our paper

Requirements

conda env create -f environment.yml

Dataset Preparation

1. Download the original dataset MVTec, PACS and CIFAR-10.

2. Generate the corrupted test set for MVTec and CIFAR-10.

python generate_corrupted_mvtec.py

python generate_corrupted_cifar10.py

Arrange data with the following structure (e.g. MVTec dataset):

Path/To/Dataset
├── mvtec
      ├── bottle
      ├── ......
├── mvtec_brightness
      ├── bottle
      ├── ......
├── mvtec_contrast
      ├── bottle
      ├── ......
├── mvtec_defocus_blur
      ├── bottle
      ├── ......
├── mvtec_gaussian_noise
      ├── bottle
      ├── ......

Modify the file path in the scripts.

Training

For the training process, please simply execute (e.g. MVTec dataset):

python train_mvtec_fico.py

Inference

For the inference process, please simply execute (e.g. MVTec dataset):

python inference_mvtec_ATTA.py

Acknowledgment

We thank the authors from ADShift for reference. We modify their code to implement FiCo.

Citation

@article{chen2024filter,
  title={Filter or Compensate: Towards Invariant Representation from Distribution Shift for Anomaly Detection},
  author={Chen, Zining and Luo, Xingshuang and Wang, Weiqiu and Zhao, Zhicheng and Su, Fei and Men, Aidong},
  journal={arXiv preprint arXiv:2412.10115},
  year={2024}
}