SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance

(AAAI 2025 Oral Presentation)

Hongyu Yan^*1, Zijun Li^*2, Kunming Luo¹, Li Lu², Ping Tan¹

¹Hong Kong University of Science and Technology, ²Sichuan University

Pretrained Models   Paper   Arxiv  

✨ News

• We open-source the 3D native generation model CraftsMan3D. Welcome to discuss the next-generation method for point cloud completion.
• We have fixed the problem of low training speed when using the DataParallel. Adding "--deterministic" in the running command to set the deterministic option for the CUDNN backend.

Introduction

This repository is the code for SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance (AAAI 2025 Oral Presentation)

SymmCompletion is a symmetry-based method for point cloud completion. It efficiently estimates point-wise local symmetry transformation to generate geometry-align partial-missing pairs and initial point clouds. Then, it leverages the geometric features of partial-missing pairs as the explicit symmetric guidance to refine initial point clouds. Qualitative and quantitative evaluations on several benchmark datasets demonstrate that SymmCompletion outperforms state-of-the-art completion networks.

Installation

git clone https://github.com/HongyuYann/SymmCompletion.git
cd SymmCompletion
conda create --name SymmCompletion python=3.11.0
conda activate SymmCompletion
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
sh extensions/install.sh

Training and Testing

1. Download datasets

   PCN dataset

   MVP dataset

   ShapeNet55/34

   KITTI ❗Note: The official link for the KITTI dataset generated by PCN is currently returning a 404 error, we provide a replacement link to download this dataset. This replacement link will be removed once the official link is functional again.

2. Replace the original data path in dataset_configs to your path

   PCN dataset:
   PARTIAL_POINTS_PATH: your path/%s/partial/%s/%s/%02d.pcd
   COMPLETE_POINTS_PATH: your path/%s/complete/%s/%s.pcd
   
   MVP dataset (16K):
   N_POINTS: 16384
   PARTIAL_POINTS_PATH: your path/mvp_%s_input.h5
   COMPLETE_POINTS_PATH: your path/mvp_%s_gt_%dpts.h5
   
   ShapeNet55/34 dataset:
   PC_PATH: your path/shapenet_pc
   
   KITTI dataset:
   CLOUD_PATH: your path/KITTI/cars/%s.pcd
   BBOX_PATH: your path/KITTI/bboxes/%s.txt
   

3. Training

   python main.py --config cfgs/PCN_models/SymmCompletion.yaml --val_freq 10 --val_interval 50 --exp_name train_pcn --deterministic
   

4. Testing

   python main.py --config cfgs/PCN_models/SymmCompletion.yaml --test --test_interval 50 --ckpt ./ckpts/PCN/ckpt-best.pth --exp_name test_pcn
   

Pretrained Models

We provide all pretrained models in Google Drive

Visualized Results

🤗 Acknowledgements

Our code is built on AnchorFormer codebase. Our work is also inspired by these projects:

• FBNet
• PointNet++
• SnowflakeNet
• PCN
• VRCNet
• PoinTr
• GRNet

📖 BibTeX

@inproceedings{yan2025symmcompletion,
    title={SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance},
    author={Yan, Hongyu and Li, Zijun and Luo, Kunming and Lu, Li and Tan, Ping},
    booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
    volume={39},
    number={9},
    pages={9094--9102},
    year={2025}
}