Installation

We use Ubuntu 22.04/24.04, Python 3.9, PyTorch 2.1.1 and CUDA 11.8 for this project. The extensions Chamfer and PointNet2 are compiled with GCC 9. The model is trained on NVIDIA RTX 3090 GPUs.

You may refer to the instructions below to set up the environment and install the dependencies.

git clone https://github.com/Rinfly/Hyper-PCN.git
cd Hyper-PCN
conda create -n hyper-pcn python=3.9
conda activate hyper-pcn
conda install pytorch==2.1.1 torchvision==0.16.1 pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt
sh extensions/install.sh

Dataset

Download the datasets from the following links:

• PCN
• MVP
• ShapeNet55/34
• KITTI

After downloading the datasets, replace the placeholder paths in cfgs/dataset_configs with the corresponding local paths.

For PCN, update PCN.yaml:

PARTIAL_POINTS_PATH: /path/to/PCN/%s/partial/%s/%s/%02d.pcd
COMPLETE_POINTS_PATH: /path/to/PCN/%s/complete/%s/%s.pcd

For MVP, update MVP.yaml:

PARTIAL_POINTS_PATH: /path/to/MVP/mvp_%s_input.h5
COMPLETE_POINTS_PATH: /path/to/MVP/mvp_%s_gt_%dpts.h5

For ShapeNet55/34, update PC_PATH in ShapeNet-55.yaml, ShapeNet-34.yaml, and ShapeNet-Unseen21.yaml:

PC_PATH: /path/to/ShapeNet55-34/shapenet_pc

For KITTI, update KITTI.yaml:

CLOUD_PATH: /path/to/KITTI/cars/%s.pcd
BBOX_PATH: /path/to/KITTI/bboxes/%s.txt

Training and Tesing

Plase refer to train.sh and test.sh for the training and testing commands.

Acknowledgement

This code is built upon PoinTr. We are also grateful for the open-source code of DeepHypergraph, GRNet, Pointnet2_PyTorch, DGCNN and SymmCompletion.

Bibtex

@inproceedings{li2026hyper,
  title={Hyper-PCN: Hypergraph-Based Point Cloud Completion via High-Order Correlation Modeling},
  author={Li, Linfei and Tan, Pei and Li, Siqi and Zou, Changqing and Gao, Yue},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={39121--39130},
  year={2026}
}