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.
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.
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.
Weikai Chen: [email protected]
Any discussions or concerns are welcomed!
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}
}