kampelmuehler / 3psdf

3-Pole Signed Distance Function (3PSDF)

This repository contains the code for "3PSDF: Three-Pole Signed Distance Function for Learning Surfaces with Arbitrary Topologies" (CVPR 2022) by Weikai Chen, Cheng Lin, Weiyang Li and Bo Yang. The project page can be found here.

Contents

1. Introduction
2. Install
3. Contacts

Introduction

3-Pole Signed Distance Function (3PSDF) is a learnable implicit representation that is capable of representing surfaces with arbitrary topologies, including open surfaces. Unlike unsiged distance functions, 3PSDF can be easily converted into mesh using the classic iso-surface extraction technique, e.g. the Marching Cubes algorithm. 3PSDF can be learned in a manner as simple as 3-way classification, which only requires a slight change for existing frameworks based on occupancy prediction.

This code repository currently contains two parts of code: (1) C++ code for computing 3PSDF of an input mesh and the corresponding sample data for network training, and (2) Python code for training single-view reconstruction using 3PSDF.

Install

Detailed instructions on installing and using the data generation and single-view reconstruction code can be found in the ReadMe in each repository: here for data generation and here for single-view reconstruction.

Contacts

Weikai Chen: [email protected]

Any discussions or concerns are welcomed!

Citation

If you find our project useful in your research, please consider citing:

@article{chen_2022_3psdf,
  title={3PSDF: Three-Pole Signed Distance Function for Learning Surfaces with Arbitrary Topologies},
  author={Chen, Weikai and Lin, Cheng and Li, Weiyang and Yang, Bo},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  month={June},
  year={2022}
}