PaCo: Parametric Point Cloud Completion

CVPR 2025

PaCo implements parametric completion, a new point cloud completion paradigm that recovers parametric primitives rather than individual points, for polygonal surface reconstruction.

🤹‍♂️ Demo

Simply click the badge below to run the demo:

🛠️ Setup

Prerequisites

Before you begin, ensure that your system has the following prerequisites installed: * Conda * CUDA Toolkit * gcc & g++

The code has been tested with Python 3.10, PyTorch 2.6.0 and CUDA 11.8.

Installation

1. Clone the repository and enter the project directory:

bash git clone https://github.com/parametric-completion/paco && cd paco

1. Install dependencies:

Create a conda environment with all required dependencies: bash . install.sh

🚀 Usage

• Download the preprocessed ABC data: to ./data/abc:

bash python ./scripts/download_data.py

• (Optional) Download pretrained weights: to ./ckpt/ckpt-best.pth:

bash python ./scripts/download_ckpt.py

🎯 Training

• Start training using one of the two parallelization:

Distributed Data Parallel (DDP):

```bash
# Replace device IDs with your own
CUDA_VISIBLE_DEVICES=0,1 ./scripts/train_ddp.sh
```

Data Parallel (DP):

```bash
# Replace device IDs with your own
CUDA_VISIBLE_DEVICES=0,1 ./scripts/train_dp.sh
```

• Monitor training progress using TensorBoard:

bash # Replace ${exp_name} with your experiment name (e.g., default) # Board typically available at http://localhost:6006 tensorboard --logdir './output/${exp_name}/tensorboard'

📊 Evaluation

• Start evaluation of the reconstruction:

bash # Default checkpoint at `./ckpt/ckpt-best.pth` CUDA_VISIBLE_DEVICES=0,1 ./scripts/test.sh

The results will be saved to ${output_dir}/evaluation.csv.

⚙️ Available configurations

```bash

Check available configurations for training

python train.py --cfg job

Check available configurations for evaluation

python test.py --cfg job ```

Alternatively, review the main configuration file: conf/config.yaml.

🚧 TODOs

• [x] Demo and pretrained weights
• [x] Dataset and evaluation script
• [x] Hugging Face model

🎓 Citation

If you use PaCo in a scientific work, please consider citing the paper:

[paper]  [supplemental]  [arxiv]  [bibtex]
bibtex @InProceedings{chen2025paco, title = {Parametric Point Cloud Completion for Polygonal Surface Reconstruction}, author = {Zhaiyu Chen and Yuqing Wang and Liangliang Nan and Xiao Xiang Zhu}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, month = {June}, year = {2025}, pages = {11749-11758} }

🙏 Acknowledgements

Part of our implementation is based on the PoinTr repository. We thank the authors for open-sourcing their great work.