Awesome

We release the code of the paper <a href="https://arxiv.org/abs/2308.11441">Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection</a> in this repository.

Reconstruction Results

ShapeNetCars

3DScenes

KITTI

Point Upsampling Results

Point Normal Estimation Results

Installation

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

• Install python Dependencies

conda create -n levelsetudf python=3.8
conda activate levelsetudf
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
pip install tqdm pyhocon==0.3.57 trimesh PyMCubes scipy point_cloud_utils==0.29.7

• Compile C++ extensions

cd extensions/chamfer_dist
python setup.py install

Quick Start

For a quick start, you can train our LevelSetUDF to reconstruct surfaces from a single point cloud as:

python run.py --gpu 0 --conf confs/object.conf --dataname demo_car --dir demo_car

• We provide the data for a demo car in the ./data folder for a quick start on LevelSetUDF.

You can find the outputs in the ./outs folder:

│outs/
├──demo_car/
│  ├── mesh
│  ├── densepoints
│  ├── normal

• The reconstructed meshes are saved in the mesh folder
• The upsampled dense point clouds are saved in the densepoints folder
• The estimated normals for the point cloud are saved in the normal folder

Use Your Own Data

We also provide the instructions for training your own data in the following.

Data

First, you should put your own data to the ./data/input folder. The datasets is organised as follows:

│data/
│── input
│   ├── (dataname).ply/xyz/npy

We support the point cloud data format of .ply, .xyz and .npy

Run

To train your own data, simply run:

python run.py --gpu 0 --conf confs/object.conf --dataname (dataname) --dir (dataname)

Notice

• For achieving better performances on point clouds of different complexity, the weights for the losses should be adjusted. For example, we provide two configs in the ./conf folder, i.e., object.conf and scene.conf. If you are reconstructing large scale scenes, the scene.conf is recomended, otherwise, the object.conf should work fine for object-level reconstructions.

• In different datasets or your own data, because of the variation in point cloud density, this hyperparameter scale has a very strong influence on the final result, which controls the distance between the query points and the point cloud. So if you want to get better results, you should adjust this parameter. We give 0.25 * np.sqrt(POINT_NUM_GT / 20000) here as a reference value, and this value can be used for most object-level reconstructions.

Related works

Please also check out the following works that inspire us a lot:

• Junsheng Zhou et al. - Learning consistency-aware unsigned distance functions progressively from raw point clouds. (NeurIPS2022)
• Baorui Ma et al. - Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces (ICML2021)
• Baorui Ma et al. - Surface Reconstruction from Point Clouds by Learning Predictive Context Priors (CVPR2022)
• Baorui Ma et al. - Reconstructing Surfaces for Sparse Point Clouds with On-Surface Priors (CVPR2022)

Citation

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

@inproceedings{zhou2023levelset,
title={Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection},
author={Zhou, Junsheng and Ma, Baorui and Li, Shujuan and Liu, Yu-Shen and Han, Zhizhong},
booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
year={2023}
}