Website: https://realiad4ad.github.io/Real-IAD_D3/

Real-IAD D³ is released for research purpose only.

If your Hugging face is registrated by your college email address, we will approve your access request directly, otherwise please send an email to [email protected] from your affiliation email address. Thank your understanding and cooperation.

A recommended application email format is:

I am writing to request access to the Real-IAD D³ for research purposes.

Name: [Your Full Name]

Affilication: [Name of your Institution/Organization]

Title：[Your Title]

Supervisor: [Name of your supervisor](This information is only necessary if you are a student)

D³M is a comprehensive framework for industrial anomaly detection, localization, and classification, leveraging a variety of deep learning-based feature extractors for 2D (RGB) and 3D (Point Cloud) data, along with flexible fusion strategies. This repository provides an implementation of the D3M framework and tools for evaluating its performance on standard and custom datasets.

The project is designed to explore and benchmark different uni-modal and multi-modal approaches for anomaly detection in industrial settings, inspired by the challenges of real-world defect identification.

📄 Citation

If you find this work useful in your research, please consider citing our associated paper:

@article{zhu2025real,
  title={Real-IAD D3: A Real-World 2D/Pseudo-3D/3D Dataset for Industrial Anomaly Detection},
  author={Zhu, Wenbing and Wang, Lidong and Zhou, Ziqing and Wang, Chengjie and Pan, Yurui and Zhang, Ruoyi and Chen, Zhuhao and Cheng, Linjie and Gao, Bin-Bin and Zhang, Jiangning and others},
  journal={arXiv preprint arXiv:2504.14221},
  year={2025}
}

✨ Features

Versatile Anomaly Detection Framework: Supports a wide range of feature extraction and fusion methodologies.
Multi-Modal Support: Handles both 2D RGB images and 3D point cloud data.
- RGB Feature Extractors: Utilizes pre-trained models like DINO (e.g., vit_base_patch8_224_dino).
- Point Cloud Feature Extractors: Incorporates models like PointMAE, PointBert, and traditional FPFH.
Flexible Fusion Strategies: Implements various early and late fusion techniques for combining multi-modal features (e.g., simple addition, dedicated fusion modules).
Multiple Method Configurations: Easily experiment with different combinations of backbones and fusion approaches via command-line arguments (e.g., DINO, Point_MAE, DINO+Point_MAE, DINO+Point_MAE+Fusion, and custom "ours" variants).
Dataset Compatibility:
- Works with standard 3D anomaly detection datasets like MVTec 3D-AD and Eyecandies.
- Easily adaptable to custom datasets (test_3d option).
Comprehensive Evaluation:
- Calculates image-level ROCAUC.
- Calculates pixel/point-level ROCAUC for localization.
- Calculates AU-PRO scores.
Memory Bank & Coreset Subsampling: Implements memory bank concepts and coreset subsampling for efficient training and inference, particularly with large feature sets.
Extensible: Designed to be easily extended with new feature extractors, fusion modules, or datasets.
Result Visualization: Option to save prediction maps for qualitative analysis.

🚀 Getting Started

Prerequisites

Python >= 3.9
PyTorch >= 2.5.1 (preferably with CUDA support for GPU acceleration)
Other dependencies as listed in requirements.txt

Installation

Clone the repository:

git clone [https://github.com/your-username/d3m.git](https://github.com/your-username/d3m.git)
cd d3m

Create a virtual environment (recommended):

python -m venv venv
source venv/bin/activate  # On Windows use `venv\Scripts\activate`

Or using Conda:

conda create -n d3m python=3.9
conda activate d3m

Install dependencies: A requirements.txt file is recommended. Create one with the following content or adapt it to your precise needs:

# requirements.txt

# Core Deep Learning & Computer Vision
torch>=2.5.1 # Install with specific CUDA version if needed, e.g., torch>=2.5.1+cu121
torchvision>=0.20.1
torchaudio>=2.5.1
numpy>=1.26.3
opencv-python>=4.11.0
Pillow>=11.0.0
matplotlib>=3.4.2
scikit-learn>=1.6.1
scikit-image>=0.24.0
einops>=0.8.1
timm>=1.0.15

# Utilities & Configuration
pyyaml>=5.4.1
tqdm>=4.67.1
easydict>=1.13 # or addict>=2.4.0
pandas>=2.2.3
h5py>=3.13.0

# Experiment Tracking & Model Hubs (Optional, uncomment if used)
# wandb>=0.19.10
# huggingface-hub>=0.29.3

# Potentially 3D-Specific (Uncomment if your project uses these)
# open3d>=0.19.0
# pointnet2-ops>=3.0.0 # May require custom compilation
# chamferdist>=1.0.3

Then install using:

pip install -r requirements.txt

Note on PyTorch: Ensure you install the PyTorch version compatible with your CUDA toolkit. Refer to the official PyTorch website for installation instructions.

💻 Usage

The main script for running experiments is likely named your_main_script_name.py (the one containing the run_3d_ads function and argparse definitions). You can configure experiments using command-line arguments.

Example Command

python your_main_script_name.py \
    --dataset_type mvtec3d \
    --dataset_path /path/to/your/mvtec3d_anomaly_detection \
    --method_name DINO+Point_MAE+Fusion \
    --rgb_backbone_name vit_base_patch8_224_dino \
    --xyz_backbone_name Point_MAE \
    --fusion_module_path /path/to/your/fusion_checkpoint.pth \
    --img_size 224 \
    --max_sample 400 \
    --coreset_eps 0.9 \
    --save_preds
    # Add other arguments as needed

Refer to the argparse section in the main script for a full list of available arguments and their descriptions.

📊 Output

The script will print tables summarizing the performance metrics for each class and method:

Image ROCAUC Results
Pixel ROCAUC Results
AU-PRO Results

If --save_preds is enabled, anomaly maps will be saved to the specified directory (default or ./pred_maps).

🙏 Acknowledgements

This framework builds upon concepts from various SOTA anomaly detection and feature extraction literature.
Utilizes awesome libraries like PyTorch, TIMM, Open3D, etc.

📞 Contact

For questions, issues, or suggestions, please open an issue in this repository.

Datasets:

Real-IAD-D3
/

Real-IAD

You need to agree to share your contact information to access this dataset