README.md

June 2, 2026 ยท View on GitHub

Hierarchical Gaussian Mixture Normalizing Flows Modeling for Unified Anomaly Detection

PyTorch implementation for ECCV2024 paper, Hierarchical Gaussian Mixture Normalizing Flows Modeling for Unified Anomaly Detection.

Installation

Install all packages with this command:

$ python3 -m pip install -U -r requirements.txt

Download Datasets

Please download MVTecAD dataset from MVTecAD dataset, BTAD dataset from BTAD dataset, MVTecAD-3D dataset from MVTecAD-3D dataset, and VisA dataset VisA dataset.

Training

  • Run code for training MVTecAD
python main.py --dataset mvtec --seed 0 --gpu 0

Normally, you can obtain the following results:

CategoryImage/Pixel AUCCategoryImage/Pixel AUCCategoryImage/Pixel AUC
Carpet1.000/0.994Bottle1.000/0.986Pill0.966/0.988
Grid0.997/0.991Cable0.970/0.959Screw0.961/0.993
Leather1.000/0.996Capsule0.988/0.992Toothbrush0.911/0.990
Tile1.000/0.961Hazelnut0.998/0.988Transistor0.977/0.913
Wood0.996/0.959Metal nut1.000/0.981Zipper0.999/0.990
Mean0.984/0.979
  • Run code for training BTAD
python main.py --dataset btad --seed 0 --gpu 0

Normally, you can obtain the following results:

CategoryImage/Pixel AUCCategoryImage/Pixel AUCCategoryImage/Pixel AUC
011.000/0.976020.859/0.973030.987/0.990
Mean0.949/0.980
  • Run code for training MVTecAD-3D
python main.py --dataset mvtec3d --seed 0 --gpu 0

Normally, you can obtain the following results:

CategoryImage/Pixel AUCCategoryImage/Pixel AUCCategoryImage/Pixel AUC
Bagel0.977/0.988Cable gland0.963/0.995Carrot0.889/0.988
Cookie0.734/0.966Dowel0.960/0.992Foam0.811/0.917
Peach0.829/0.994Potato0.690/0.950Rope0.976/0.992
Tire0.876/0.986Mean0.871/0.977
  • Run code for training VisA
python main.py --dataset visa --seed 0 --gpu 0

Normally, you can obtain the following results:

CategoryImage/Pixel AUCCategoryImage/Pixel AUCCategoryImage/Pixel AUC
Candle0.988/0.995Capsules0.956/0.990Cashew0.910/0.991
Chewinggum0.999/0.996Fryum0.984/0.949Macaroni10.991/0.998
Macaroni20.926/0.997Pcb10.976/0.995Pcb20.956/0.983
Pcb30.986/0.994Pcb40.979/0.987Pipe fyrum0.996/0.993
Mean0.971/0.989
  • Run code for training Union dataset (combined by MVTecAD, BTAD, MVTecAD-3D, and VisA)
python main.py --dataset union --seed 0 --gpu 0

We also report the detailed results on the Union dataset as follows:

CategoryImage/Pixel AUCCategoryImage/Pixel AUCCategoryImage/Pixel AUC
Bottle1.000/0.982Cable0.951/0.860Capsule0.934/0.990
Carpet1.000/0.993Grid0.986/0.983Hazelnut1.000/0.985
Leather1.000/0.995Metal nut0.997/0.981Pill0.969/0.984
Screw0.812/0.986Tile0.999/0.936Toothbrush0.961/0.992
Transistor0.996/0.901Wood0.994/0.957Zipper0.999/0.992
010.997/0.974020.838/0.969030.995/0.997
Bagel0.983/0.991Cable gland0.886/0.990Carrot0.815/0.990
Cookie0.792/0.972Dowel0.896/0.978Foam0.798/0.913
Peach0.856/0.993Potato0.625/0.958Rope0.929/0.994
Tire0.835/0.965
Candle0.989/0.996Capsules0.939/0.975Cashew0.928/0.987
Chewinggum0.996/0.996Fryum0.976/0.938Macaroni10.975/0.997
Macaroni20.903/0.995Pcb10.964/0.992Pcb20.966/0.972
Pcb30.964/0.990Pcb40.981/0.981Pipe fyrum0.991/0.992
Mean0.935/0.975

Note: You need to set the root directory of your dataset in the main.py by setting args.data_path. For Union dataset, the dataset path can be set in the datasets/union.py script.

Citation

If you find this repository useful, please consider citing our work:

@article{HGAD,
      title={Hierarchical Gaussian Mixture Normalizing Flows Modeling for Unified Anomaly Detection}, 
      author={Xincheng Yao and Ruoqi Li and Zefeng Qian and Lu Wang and Chongyang Zhang},
      year={2024},
      booktitle={European Conference on Computer Vision 2024},
      url={https://arxiv.org/abs/2403.13349},
      primaryClass={cs.CV}
}

If you are interested in our work, you can also see our other works: BGAD (CVPR2023), PMAD (AAAI2023), FOD (ICCV2023), ResAD (NeurIPS2024), ADPretrain (NeurIPS2025), MMR-AD (CVPR2026), ResAD++ (IJCV2026). Or, you can follow our github page xcyao00.