Anomaly-ShapeNet: A Synthetic Dataset of Point Cloud Anomaly Detection

October 17, 2025 · View on GitHub

Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, Yingna Wu

🧩Works based on Anomaly-ShapeNet and IMRNet

Mulsen-AD-3D: We propose a large scale 3D AD dataset following Anomaly-ShapeNet and Real3D-AD (CVPR2025)!

PASDF: We propose a simple and effective 3D-AD algorithm based on IMR-Net (ICCV2025)!

R3DAD: Reconstruction based 3D-AD codebase based on IMR-Net (ECCV2024).

Introduction

This project provides a synthetic dataset for point cloud anomaly detection. We provide 3D point clouds and meshs for training and testing 3D anomaly detection algorithms. Note that we expanded from the original 40 classes to 50 classes and put the new 10 classes in the ‘new’ folders.Our paper can be found here.[Paper]

Data description

overview of Anomaly-ShapeNet

Anomaly-ShapeNet comprises a total of 1600 samples which are distributed across 40(+10) distinct categories. There are six kinds of anomalies, including bulge, concavity, crack, holes, and broken. All the obj or pcd file have been watertighted before to obtain a more smooth surface.

Datasets	Category numbers	Anomaly types	Train num(1cls)	Test num(1cls)	Point num	Anomaly portion
Anomaly-ShapeNet	50	6	4	15~24	8K~30K	1%~7%

Download

To download our Anomaly-ShapeNet dataset (Dataset for training and evaluation, pcd&obj format), click Anomaly-ShapeNet.zip(baidu disk: case) or (hugging face)(https://huggingface.co/datasets/Chopper233/Anomaly-ShapeNet)

Data preparation

Download Anomaly-ShapeNet-v2.zip and extract into ./dataset/pcd

pcd
├── ashtray0
    ├── train
        ├── *template.pcd
        ...
    ├── test
        ├── 1_bulge.pcd
        ├── 2_concavity.pcd
        ...
    ├── GT
        ├── 1_bulge.txt
        ├── 2_sink.txt
        ... 
├── bag0
...

Benchmark

Our benchmark for Anomaly ShapaNet and Real3D-AD are listed here Benchmark. Some metrics like P-AP for Anomaly-ShapeNet is ignored because they are not suitable for our dataset.

Citation

Please cite the following paper if this dataset helps your project:

@InProceedings{Li_2024_CVPR,
    author    = {Li, Wenqiao and Xu, Xiaohao and Gu, Yao and Zheng, Bozhong and Gao, Shenghua and Wu, Yingna},
    title     = {Towards Scalable 3D Anomaly Detection and Localization: A Benchmark via 3D Anomaly Synthesis and A Self-Supervised Learning Network},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {22207-22216}
}

Thanks

Our dataset and code are built on ShapeNet and PointMAE, thanks for their excellent works!

License

The dataset is released under the CC BY 4.0 license.