Uni3DAD

This is not the latest code. I will update the latest version...

This is the implementation of Uni-3DAD: GAN-Inversion Aided Universal 3D Anomaly Detection on Model-free Products.

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

@article{LIU2025126665,
title = {Uni-3DAD: Gan-inversion aided universal 3D anomaly detection on model-free products},
journal = {Expert Systems with Applications},
volume = {272},
pages = {126665},
year = {2025},
issn = {0957-4174},
doi = {https://doi.org/10.1016/j.eswa.2025.126665},
url = {https://www.sciencedirect.com/science/article/pii/S0957417425002878},
author = {Jiayu Liu and Shancong Mou and Nathan Gaw and Yinan Wang},
keywords = {Unsupervised learning, 3D point clouds, Anomaly detection}
}

How to use code?

1. Environment:

Linux: 20.04
Python: 3.8.15
Pytorch: 1.13.1
CUDA: 11.7

2. Clone the repo:

git clone https://github.com/JiayuLiu666/Uni3DAD.git

3. Pip necessary packages:

Since this repo is built on 3D-ADS (https://github.com/eliahuhorwitz/3D-ADS), M3DM (https://github.com/nomewang/M3DM), Shape-Inversion (https://github.com/junzhezhang/shape-inversion.git), and Shape-guided (https://github.com/jayliu0313/Shape-Guided.git), please refer to their Github pages for the necessary packages and pre-trained models. Thanks for their contributions.

4. Dataset:

We use MVTec 3D-AD as our dataset (https://www.mvtec.com/company/research/datasets/mvtec-3d-ad). Please refer to 3D-ADS, M3DM, and Shape-guided for the data preprocessing.
(We also create our own dataset for missing parts detection based on MVTec 3D-AD; if you need it, please contact us.)

data_process.ipynb is the code that creates the validation dataset.

5. Pre-trained models:

You can use

train_dist.py 

to train your own GAN models for each category. We also provide the pre-trained models:

After training is finished, you can use

visual_dist.py

to check the results of training.

The GAN models and necessary feature-extractors models should be saved like this structure:

├──Checkpoints
│   ├── best_ckpt
│   │   ├── ckpt_00601.pth
│   ├── ...
│   ├── pointMAE_pretrain.pth
│   ├── ...
├── Common
│   ├── ...
│   ├── ...
├── pretrain_checkpoints
│   ├── bagel.ckpt
│   ├── ...
└── README.md

Run the code

You can use

python runner.py --METHOD_NAME BTF+GAN --saved_training "YOUR Directory" ...

to run the code. Please refer to config.py in the Generation to change the parameters.

Dataset

https://drive.google.com/file/d/1FozkWrj1Y7MBX_UaA6G8Eaiwn1aHQIqr/view?usp=drive_link