Towards High-Resolution 3D Anomaly Detection: A Scalable Dataset and Real-Time Framework for Subtle Industrial Defects

November 19, 2025 · View on GitHub

🌐 Hugging Face Dataset

📚 Paper • 🏠 Homepage
by , Yuqi Cheng*, Yihan Sun*, Hui Zhang Weiming Shen, Yunkang Cao

🚀 Updates

We're committed to open science! Here's our progress:

  • 2025/11/19: More comprehensive, realistic, and flexible anomaly synthesis framework will be open-sourced in our upcoming work!
  • 2025/11/19: 🎉 Code of Simple3D is available.
  • 2025/11/08: 🎉 Our paper has been accepted by AAAI 2026 (Oral).
  • 2025/07/10: 📄 Paper released on ArXiv.
  • 2025/07/08: 🌐 Dataset homepage launched.

📊 Introduction

In industrial point cloud analysis, detecting subtle anomalies demands high-resolution spatial data, yet prevailing benchmarks emphasize low-resolution inputs. To address this disparity, we propose a scalable pipeline for generating realistic and subtle 3D anomalies. Employing this pipeline, we developed MiniShift, the inaugural high-resolution 3D anomaly detection dataset, encompassing 2,577 point clouds, each with 500,000 points and anomalies occupying less than 1% of the total. We further introduce Simple3D, an efficient framework integrating Multi-scale Neighborhood Descriptors (MSND) and Local Feature Spatial Aggregation (LFSA) to capture intricate geometric details with minimal computational overhead, achieving real-time inference exceeding 20 fps. Extensive evaluations on MiniShift and established benchmarks demonstrate that Simple3D surpasses state-of-the-art methods in both accuracy and speed, highlighting the pivotal role of high-resolution data and effective feature aggregation in advancing practical 3D anomaly detection.

🔍 Overview of MiniShift

12 Categories and 4 defect types

Anchor-Guided Geometric Anomaly Synthesis

Download

You are welcome to try our dataset: Hugging Face Dataset and Baidi Drive

Simple3D

🛠️ Getting Started

Clone and install dependencies:

git clone https://github.com/hustCYQ/MiniShift-Simple3D.git && cd MiniShift-Simple3D  
conda create --name Simple3D_env python=3.8 -y  
conda activate Simple3D_env  

conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
pip install tifffile open3d-cpu
pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl
git clone https://github.com/erikwijmans/Pointnet2_PyTorch.git
cd Pointnet2_PyTorch
pip install -r requirements.txt
pip install -e .
cd ..

🚀 Train & Evaluate

# Example: Run in MiniShift 
--level : ALL/easy/medium/hard
--vis_save: Save anomaly scores
python ./main.py --dataset minishift --num_group 4096 --group_size 128  --max_nn 40 --use_LFSA True --use_MSND True --num_MSND 2 --expname MiniShift_ALL --level ALL --vis_save True

# Evaluate on other datasets.72

python ./main.py --dataset real --num_group 4096 --group_size 128  --max_nn 40 --use_LFSA True --use_MSND True --expname real
python ./main.py --dataset shapenet --num_group 4096 --group_size 128  --max_nn 40 --use_LFSA True --use_MSND True --expname shapenet
python ./main.py --dataset mulsen --num_group 4096 --group_size 128  --max_nn 40 --use_LFSA True --use_MSND True --expname mulsen

You can selectively reduce group_size and max_nn to balance efficiency and accuracy.

We also simply reproduced many descriptors for selection, such as SHOT, CVFH, Spin, et al.

--feature : FPFH/shape_context/CVFH/NARF/Spin/USC/SHOT
python ./main.py --dataset minishift --num_group 4096 --group_size 128  --max_nn 40 --use_LFSA True --num_MSND True --expname MiniShift_ALL --level ALL --feature SHOT

Render the anomaly detection results to a .mp4 file:

python render_video.py --input_paths "./vis-results/MulSen_AD/capsule/test/crack/0.txt" "./vis-results/MulSen_AD/capsule/gt/crack/0.txt" --output_path "capsule_crack.mp4"

📊 Main Results

1. Performance on MiniShift

2. Performance on Real3D-AD, Anomaly-ShapeNet, and MulSenAD

🙏 Acknowledgements

Grateful to these projects for inspiration:

📖 Citation

If our work aids your research, please cite:

@article{MiniShift_Simple3D,  
  title={Towards High-Resolution 3D Anomaly Detection: A Scalable Dataset and Real-Time Framework for Subtle Industrial Defects},
  author={Cheng, Yuqi and Sun, Yihan and Zhang, Hui and Shen, Weiming and Cao, Yunkang},
  journal={arXiv preprint arXiv:2507.07435},
  year={2025}
}

Contact

If you have any questions about our work, please do not hesitate to contact yuqicheng@hust.edu.cn.