Constrained Adaptive Projection with Pretrained Features for Anomaly Detection(IJCAI 2022)

pytorch implementation of paper "Constrained Adaptive Projection with Pretrained Features for Anomaly Detection" （CAP）

arxiv paper address: https://arxiv.org/abs/2112.02597

official ijcai paper address: https://www.ijcai.org/proceedings/2022/0286.pdf

1. Requirements

Currently, requires following packages

• python 3.9.5
• torch 1.9.0
• CUDA 11.1
• torchvision 0.10
• faiss 1.7.1
• scikit-learn 0.24.2

2. Framework

3. Experiment

quickly run

For cifar10

sh cifar10.sh

For cifar100, exchange --dataset cifar100

For mvTec, please download mvTec dataset and exchange hyperparameters corresponding to appendix file.

run a specific class

python main.py --dataset <dataset> --normal_class <normal-class> --regular <constrained lambda>

Citation

If you find this project useful in your research, please consider cite:

@article{gui2021constrained,
  title={Constrained Adaptive Projection with Pretrained Features for Anomaly Detection},
  author={Gui, Xingtai and Wu, Di and Chang, Yang and Fan, Shicai},
  journal={arXiv preprint arXiv:2112.02597},
  year={2021}
}