BridgeNet: A Unified Multimodal Framework for Bridging 2D and 3D Industrial Anomaly Detection

December 29, 2025 · View on GitHub

An Xiang*, Zixuan Huang*, Xitong Gao*, Kejiang Ye†, Cheng-zhong Xu (* Equal contribution; † Corresponding authors)

Our paper has been accepted by ACM MM 2025 [Paper].

Quick Start

1. Environment Installation

Prerequisites

  • Python 3.9
  • Conda package manager

Option A: Automatic Installation

# Clone the repository
git clone https://github.com/Bridgenet-3D/BridgeNet.git
cd BridgeNet

# Create conda environment (environment name is 'bridgenet')
conda env create -f environment.yml

# Activate environment
conda activate bridgenet

Option B: Manual Installation

# Create new conda environment
conda create -n bridgenet python=3.9
conda activate bridgenet

pip install torch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 --index-url https://download.pytorch.org/whl/cu118

# Install other dependencies
pip install ...

2. Dataset Structure

The dataset should be organized in the following structure:

mvtec_process/
├── cookie/
│   ├── train/
│   │   └── good/
│   │       ├── 000.png
│   │       ├── 001.png
│   │       └── ...
│   ├── test/
│   │   ├── good/
│   │   │   ├── 000.png
│   │   │   └── ...
│   │   ├── crack/
│   │   │   ├── 000.png
│   │   │   └── ...
│   │   ├── contamination/
│   │   │   ├── 000.png
│   │   │   └── ...
│   │   └── ...
│   └── ground_truth/
│       ├── crack/
│       │   ├── 000.png
│       │   └── ...
│       ├── contamination/
│       │   ├── 000.png
│       │   └── ...
│       └── ...
├── dowel/
│   ├── train/
│   │   └── good/
│   │       ├── 000.png
│   │       ├── 001.png
│   │       └── ...
│   ├── test/
│   │   ├── good/
│   │   │   ├── 000.png
│   │   │   └── ...
│   │   ├── bent/
│   │   │   ├── 000.png
│   │   │   └── ...
│   │   ├── cut/
│   │   │   ├── 000.png
│   │   │   └── ...
│   │   └── ...
│   └── ground_truth/
│       ├── bent/
│       │   ├── 000.png
│       │   └── ...
│       ├── cut/
│       │   ├── 000.png
│       │   └── ...
│       └── ...
├── ...
│      
└── Depth

Each category (e.g., cookie, dowel) contains:

  • train/good/: Normal training samples
  • test/good/: Normal test samples
  • test/<anomaly>/: Anomalous test samples
  • ground_truth/<anomaly>/: Ground truth masks for anomalies

The dataset formatted [mvtec3d_formatted]

Original dataset [mvtec3d]

Texture anomaly dataset [dtd]

3. Running BridgeNet

We provide a simple shell script to run BridgeNet on the MVTec-3D dataset:

cd shell
bash run-mvtec.sh

4. Results

Results will be saved in the results/ directory with the following structure: /root/3D/BridgeNet-main_/results/models/backbone_0/mvtec3d_foam

results/
├── models/
│   └── backbone_0/
│       └── class_name/
├── eval
│   └── class_name/
└── training/            
    └── class_name/

Citation

If you use BridgeNet in your research, please cite:

@inproceedings{xiang2025bridgenet,
  title={BridgeNet: A Unified Multimodal Framework for Bridging 2D and 3D Industrial Anomaly Detection},
  author={Xiang, An and Huang, Zixuan and Gao, Xitong and Ye, Kejiang and Xu, Cheng-zhong},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={1579--1587},
  year={2025}
}

Acknowledgments

We would like to thank the following repositories for their valuable contributions and support:

These open-source projects have been instrumental in advancing the field of anomaly detection.