This repository is an official implementation of the following paper:

Point Cloud Classification Using Content-based Transformer via Clustering in Feature Space

Yahui Liu, Bin Tian, Yisheng Lv, Lingxi Li, Feiyue Wang

Accepted for publication in the IEEE/CAA Journal of Automatica Sinica

# clone this repo
git clone https://github.com/yahuiliu99/PointConT.git
cd PointConT

# create a conda env
conda create -n pointcont -y python=3.7 numpy=1.20 numba
conda activate pointcont

# install PyTorch and libs (refer to requirements.txt)
# please install compatible PyTorch and CUDA versions
conda install -y pytorch=1.10.1 torchvision cudatoolkit=11.1 -c pytorch -c nvidia
pip install hydra-core==1.1 h5py scikit-learn einops tqdm warmup-scheduler deepspeed tensorboard  

# install the pointnet++ library cuda extensions
pip install pointnet2_ops_lib/.

When you first run the command for training, the datasets will be automatically downloaded and saved in data/.

• ModelNet40 -->data/modelnet40_ply_hdf5_2048/
• ScanObjectNN -->data/h5_files/

Alternatively, you can manually download the official data (ModelNet40 | ScanObjectNN) in any path, and create a symbolic link to your datasets folder.

mkdir data
ln -s /path/to/your/data/folder data/

You can modify settings in config/cls.yaml.

Make sure the eval is set to False.

We support wandb for collecting results online. Just set wandb.use_wandb=True if use wandb. Please check the official wandb doc for more details.

• Classification on ModelNet40

  python main_cls.py db=modelnet40
  

• Classification on ScanObjectNN

  python main_cls.py db=scanobjectnn
  

config/cls.yaml will be automatically loaded when you run the command.

To evaluate a trained-model, please set eval=True in config/cls.yaml and run python main_cls.py db=${dataset}

Or you can override values in the loaded config from the command line:

python main_cls.py db=${dataset} eval=True

• Mitsuba

Please refer to the following github repository for point cloud rendering code: PointFlowRenderer

* denotes method evaluated with voting strategy

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

@article{Liu2023PointConT,
    author = {Liu, Yahui and Tian, Bin and Lv, Yisheng and Li, Lingxi and Wang, Fei-Yue},
    title = {Point Cloud Classification Using Content-based Transformer via Clustering in Feature Space},
    journal = {IEEE/CAA Journal of Automatica Sinica}, 
    year={2023},
    volume={10},
    number={8},
    pages={1-9},
    doi={10.1109/JAS.2023.123432}
}

Our code is mainly based on the following open-source projects. Many thanks to the authors for their wonderful works.

PointNet2, Point-Transformers, DGCNN, CurveNet, PointMLP, PAConv, PointNeXt