MC3D-AD

January 18, 2026 · View on GitHub

Official PyTorch Implementation of MC3D-AD: A Unified Geometry-aware Reconstruction Model for Multi-category 3D Anomaly Detection. Accepted by IJCAI 2025.

Image text

1. Quick Start

1.1 Requirements

  • Conda virtual environment creation
conda create -n MC3D-AD python=3.8
conda activate MC3D-AD
pip install -r requirements.txt
pip install "git+https://github.com/erikwijmans/Pointnet2_PyTorch.git#egg=pointnet2_ops&subdirectory=pointnet2_ops_lib"
pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl
  • Download feature extractor pre-trained weights. Point-MAE pre-trained weights can be downloaded here. After downloading, place the modelnet_8k.pth file in the ./pretrain_ckp directory

1.2 Real3D-AD

  • How to get the Real3D-AD dataset. Download the Real3D-AD dataset from here. After downloading, unzip it to the directory you specified.

  • Real3D Downsampling. To simplify the training, you first need to downsampling the training set of Real3D. You can use the following command: python downsample_pcd.py --real3d_path <Path/to/your/Real3D-AD-PCD> After the above operations are completed, the real3D_down directory will be created in the same directory layer as Real3D-AD-PCD.

  • cd the experiment directory by running cd ./experiments/real3d/.

  • Modify the config.yaml. Please set the dataset.data_dir and net.data_dir of config.yaml to load the absolute path of your real3D_down.

  • Train or eval by running:

    sh train.sh 1 #GPU_ID or sh eval.sh #1 #GPU_ID.

    Note: Multi-gpu training is not currently supported. During eval, please set the saver.load_path of config.yaml to load the checkpoints.

1.3 Anomaly-ShapeNet

  • How to get the Anomaly-ShapeNet dataset. Download the Anomaly-ShapeNet dataset from here. After downloading, unzip it to the directory you specified. Take out Anomaly-ShapeNet-v2/dataset/pcd and organize it into the following format:
Anomaly-shapeNet
├── ashtray0
    ├── train
        ├── *template.pcd
        ...
    ├── test
        ├── 1_bulge.pcd
        ├── 2_concavity.pcd
        ...
    ├── GT
        ├── 1_bulge.txt
        ├── 2_sink.txt
        ... 
├── bag0
...

  • Anomaly-ShapeNet dataset setting. Please change the name of the pcd directory to Anomaly-shapeNet and fill the absolute path of the Anomaly-shapeNet into dataset.data_dir and net.data_dir in ./experiments/Anomaly-ShapeNet/config.yaml.

  • cd the experiment directory by running cd ./experiments/Anomaly-ShapeNet/.

  • Train or eval by running:

    sh train.sh 1 #GPU_ID or sh eval.sh 1 #GPU_ID.

    Note: During eval, please set config.saver.load_path to load the checkpoints. If you have any issues, please go to https://github.com/jiayi-art/MC3D-AD. The appendix for MC3D-AD has been uploaded to this repository.

Acknowledgement

Our work is inspired by UniAD , Reg3D-AD and IMRNet, thanks their extraordinary works!