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:

Category	Image/Pixel AUC	Category	Image/Pixel AUC	Category	Image/Pixel AUC
Carpet	1.000/0.994	Bottle	1.000/0.986	Pill	0.966/0.988
Grid	0.997/0.991	Cable	0.970/0.959	Screw	0.961/0.993
Leather	1.000/0.996	Capsule	0.988/0.992	Toothbrush	0.911/0.990
Tile	1.000/0.961	Hazelnut	0.998/0.988	Transistor	0.977/0.913
Wood	0.996/0.959	Metal nut	1.000/0.981	Zipper	0.999/0.990
Mean	0.984/0.979

• Run code for training BTAD

python main.py --dataset btad --seed 0 --gpu 0

Normally, you can obtain the following results:

Category	Image/Pixel AUC	Category	Image/Pixel AUC	Category	Image/Pixel AUC
01	1.000/0.976	02	0.859/0.973	03	0.987/0.990
Mean	0.949/0.980

• Run code for training MVTecAD-3D

python main.py --dataset mvtec3d --seed 0 --gpu 0

Normally, you can obtain the following results:

Category	Image/Pixel AUC	Category	Image/Pixel AUC	Category	Image/Pixel AUC
Bagel	0.977/0.988	Cable gland	0.963/0.995	Carrot	0.889/0.988
Cookie	0.734/0.966	Dowel	0.960/0.992	Foam	0.811/0.917
Peach	0.829/0.994	Potato	0.690/0.950	Rope	0.976/0.992
Tire	0.876/0.986	Mean	0.871/0.977

• Run code for training VisA

python main.py --dataset visa --seed 0 --gpu 0

Normally, you can obtain the following results:

Category	Image/Pixel AUC	Category	Image/Pixel AUC	Category	Image/Pixel AUC
Candle	0.988/0.995	Capsules	0.956/0.990	Cashew	0.910/0.991
Chewinggum	0.999/0.996	Fryum	0.984/0.949	Macaroni1	0.991/0.998
Macaroni2	0.926/0.997	Pcb1	0.976/0.995	Pcb2	0.956/0.983
Pcb3	0.986/0.994	Pcb4	0.979/0.987	Pipe fyrum	0.996/0.993
Mean	0.971/0.989

• 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:

Category	Image/Pixel AUC	Category	Image/Pixel AUC	Category	Image/Pixel AUC
Bottle	1.000/0.982	Cable	0.951/0.860	Capsule	0.934/0.990
Carpet	1.000/0.993	Grid	0.986/0.983	Hazelnut	1.000/0.985
Leather	1.000/0.995	Metal nut	0.997/0.981	Pill	0.969/0.984
Screw	0.812/0.986	Tile	0.999/0.936	Toothbrush	0.961/0.992
Transistor	0.996/0.901	Wood	0.994/0.957	Zipper	0.999/0.992
01	0.997/0.974	02	0.838/0.969	03	0.995/0.997
Bagel	0.983/0.991	Cable gland	0.886/0.990	Carrot	0.815/0.990
Cookie	0.792/0.972	Dowel	0.896/0.978	Foam	0.798/0.913
Peach	0.856/0.993	Potato	0.625/0.958	Rope	0.929/0.994
Tire	0.835/0.965
Candle	0.989/0.996	Capsules	0.939/0.975	Cashew	0.928/0.987
Chewinggum	0.996/0.996	Fryum	0.976/0.938	Macaroni1	0.975/0.997
Macaroni2	0.903/0.995	Pcb1	0.964/0.992	Pcb2	0.966/0.972
Pcb3	0.964/0.990	Pcb4	0.981/0.981	Pipe fyrum	0.991/0.992
Mean	0.935/0.975

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}
}