Inter-Realization Channels (InReaCh): Unsupervised Anomaly Detection in Images Beyond One-Class Classification

This repository contains the implementation of InReaCh, a fully unsupervised method for anomaly detection and localization in images. The method was introduced in ICCV 2023 and achieves state-of-the-art performance in unsupervised anomaly detection tasks.

Also check out the newer Online-InReaCh for better performance, tolerance to non-stationary distributions and online adaptation! Online InReaCh code can be found here!

Overview

InReaCh leverages the concept of Inter-Realization Channels to extract high-confidence nominal patches from training data by associating them across image realizations. These nominal patches are used to build a model for detecting and localizing anomalies in test images. Unlike supervised one-class classification methods, InReaCh is fully unsupervised, simplifying dataset creation and broadening its applicability.

Key features of InReaCh:

• Unsupervised anomaly detection: No need for labeled training data.
• High precision: Achieves 99.9% precision in extracting nominal patches from the MVTec AD dataset.
• Robust performance: Maintains high accuracy even with up to 40% corrupted training data.
• Competitive results: Achieves 0.968 AUROC in localization and 0.923 AUROC in detection.

Tables and Figures

Repository Structure

• InReaCh.py: Main implementation of the InReaCh algorithm, including training and testing pipelines.
• FeatureDescriptors.py: Code for feature extraction and descriptor generation.
• mvtec_loader.py: Utilities for loading and preprocessing the MVTec AD dataset.
• utils.py: Helper functions for alignment, seeding, and positional tests.
• model.py: Code for loading and configuring the Wide ResNet-50 model for feature extraction.
• requirements.txt: Python dependencies for the project.
• conda_inreach.yml: Conda environment configuration file.
• images/: Contains visualizations of the method, results, and qualitative examples.
• data/: Placeholder for dataset-related files.

Installation

Using Conda

1. Clone the repository:

   git clone https://github.com/your-repo/InReaCh.git
   cd InReaCh

2. Create the Conda environment:

   conda env create -f conda_inreach.yml

3. Activate the environment:

   conda activate inreach

Using Pip

1. Install the required Python packages:

   pip install -r requirements.txt

Usage

Dataset Preparation

Download the MVTec AD dataset and place it in the data/ directory. The dataset should follow the structure expected by the mvtec_loader.py script.

Running the Code

1. Train and test the InReaCh model:

   python InReaCh.py

2. Modify parameters such as class_names, assoc_depth, and max_channel_std in the InReaCh.py file to customize the training and testing process.

Visualizing Results

The repository includes visualizations of the method and results in the images/ directory:

• Method.png: Overview of the InReaCh methodology.
• Qualitative.png: Qualitative results of anomaly detection.
• Results.png: Quantitative results and performance metrics.

Dependencies

The project requires the following key libraries:

• Python 3.12
• PyTorch 2.6.0
• torchvision 0.21.0
• OpenCV 4.11.0
• FAISS (GPU version)
• NumPy, SciPy, scikit-learn, tqdm

For a complete list of dependencies, refer to the conda_inreach.yml or requirements.txt file.

Citation

If you use this code in your research, please cite the following paper:

@misc{mcintoshInReaCh,
      title={Inter-Realization Channels: Unsupervised Anomaly Detection in Images Beyond One-Class Classification},
      author={Declan McIntosh and Alexandra Branzan Albu},
      year={2023},
      booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
      month     = {October}
}

Acknowledgments

This implementation builds upon the concepts introduced in the ICCV 2023 paper. Special thanks to the authors and contributors for their work