Multi-Flow: Multi-View-Enriched Normalizing Flows for Industrial Anomaly Detection

This is the official code to the paper "Multi-Flow: Multi-View-Enriched Normalizing Flows for Industrial Anomaly Detection" by Mathis Kruse and Bodo Rosenhahn which was accepted to the CVPR 2025 Workshop VAND 3.0.

Abstract

With more well-performing anomaly detection methods proposed, many of the single-view tasks have been solved to a relatively good degree. However, real-world production scenarios often involve complex industrial products, whose properties may not be fully captured by one single image. While normalizing flow based approaches already work well in single-camera scenarios, they currently do not make use of the priors in multi-view data. We aim to bridge this gap by using these flow-based models as a strong foundation and propose Multi-Flow, a novel multi-view anomaly detection method. Multi-Flow makes use of a novel multi-view architecture, whose exact likelihood estimation is enhanced by fusing information across different views. For this, we propose a new cross-view message-passing scheme, letting information flow between neighboring views. We empirically validate it on the real-world multi-view data set Real-IAD and reach a new state-of-the-art, surpassing current baselines in both image-wise and sample-wise anomaly detection tasks.

Installation Setup

To reproduce our results, please clone this repository. We recommend creating a new conda environment and installing all necessary modules. This code has been written and only tested on a Linux machine running openSUSE.

# (1) Cloning the repository
git clone git@github.com:m-kruse98/Multi-Flow.git
cd Multi-Flow

# (2) Create the environment with necessary CUDA & PyTorch frameworks
conda env create --file environment.yml 
conda activate multiflow

You will also need the necessary background segmentation models from the MVANet codebase, which may also found here, and put them into the "models" directory. Alternatively just run the code below to directly download the necessary checkpoints.

# (3) Make sure steps (1)-(2) are completed. Download the model checkpoints from the MVANet repository
cd models
# Download MVANet Segmentation model checkpoint (~387 MB)
gdown https://drive.google.com/uc\?id\=1_gabQXOF03MfXnf3EWDK1d_8wKiOemOv
# Download Swin Transformer Checkpoint (~365 MB)
gdown https://drive.google.com/uc\?id\=1-Zi_DtCT8oC2UAZpB3_XoFOIxIweIAyk
cd ..

Additonally run the following script, to flip some of the camera views in Real-IAD, to force multi-view consistency across images. ATTENTION: This will change the images in the dataset IN-PLACE on your disk! This is not strictly required, but may improve performance.

# (4) (OPTIONAL) Make sure the paths are correctly set in flip_images.py
python flip_images.py

Running the code

Make sure that all steps from the installation (optionally including flip_images.py) are done.

Feature & Mask Extraction

Extract features and foreground masks for a desired class (e.g. 'audiojack') using

python preprocess.py -c audiojack

This needs to be run before any anomaly detection. Important: Make sure the paths in 'config.py' point to your Real-IAD data set (including the 'realiad_jsons' directory from 'realiad_jsons.zip'). Features will be extracted to the new directory 'tmp/', with masks being saved to 'tmp/masks/'.

Note: Consider parallelizing this extraction within a (Slurm-)Cluster as it may take quite some time for each class (~30-60 minutes or more possible).

Running Anomaly Detection

Validate that the features and masks were correctly extracted to the respective directories, and that 'config.py' links to them via the 'feature_dir'. Then, to train and evaluate Multi-Flow, run

python train.py -c audiojack

Different options can be found using python train.py --help. Hyperparameters may mostly be tweaked using the 'config.py' file.

BibTeX - Citation

@InProceedings{Kruse_2025_CVPR,
    author    = {Kruse, Mathis and Rosenhahn, Bodo},
    title     = {Multi-Flow: Multi-View-Enriched Normalizing Flows for Industrial Anomaly Detection},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops},
    month     = {June},
    year      = {2025},
    pages     = {3933-3944}
}

Credits

Credits go to the following code bases. Check out their amazing work:

• ADEval (optimized AUROC/AUPRO routines)
• CS-Flow (anomaly detection and normalizing flow code)
• MVANet (pre-trained segmentation networks)
• Real-IAD (data loading routines)

Acknowledgements

This work was supported by the Federal Ministry of Education and Research (BMBF), Germany, under the AI service center KISSKI (grant no. 01IS22093C), the MWK of Lower Sachsony within Hybrint (VWZN4219), the Deutsche Forschungsgemeinschaft (DFG) under Germany’s Excellence Strategy within the Cluster of Excellence PhoenixD (EXC2122), the European Union under grant agreement no. 101136006 – XTREME.

License

This project is licensed under the MIT license. However, do take note of the licenses of the credited repositories ADEval, CS-Flow, MVANet and Real-IAD.