Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces (ICML 2021)

This repository contains the official pytorch version code for the paper. Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces.

You can find detailed usage instructions for training your own models or running our provided demo below.

If you find our code or paper useful, please consider citing

@inproceedings{NeuralPull,
    title = {Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces},
    author = {Baorui, Ma and Zhizhong, Han and Yu-Shen, Liu and Matthias, Zwicker},
    booktitle = {International Conference on Machine Learning (ICML)},
    year = {2021}
}

Tensorflow Version

This repository contains the official pytorch version code for Neural-Pull. If you are more accessible to the tensorflow code, please use tensorflow repository and star it, thanks.

Surface Reconstruction Demo

Installation:

Our code is implemented in Python 3.8, PyTorch 1.11.0 and CUDA 11.3.

• Install python Dependencies

conda create -n npull python=3.8
conda activate npull
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
pip install tqdm pyhocon==0.3.57 trimesh PyMCubes scipy

Data:

Surface Reconstruction from Point Cloud.

We provide the demo data in data/gargoyle.ply. If you want to reconstruct your own data, please:

• Put your point cloud data on ./data.
• Note that we support the point cloud data format of .ply and .xyz.

Usage:

python run.py --gpu 0 --conf confs/npull.conf --dataname gargoyle --dir gargoyle

You can find the generated mesh and the log in ./outs.

License

This project is open sourced under MIT license.