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.

1. Introduction
2. Install
3. Usage
4. Contacts

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.

The folder data_generation contains the C++ code for generating 3PSDF samples.

This repository is self-contained -- you do not need to install any external libraries. Simply run the following lines to compile the code:

cd data_generation
mkdir build
cd build
cmake ..
make -j8

The code is only tested on Ubuntu 16.04 and 18.04.

TODO

After compilation, binaries named gen_3psdf_samples and batch_generate will be generated.

(1) gen_3psdf_samples: the binary generates 3PSDF sampling points using octree-based sampling.

Simply run ./gen_3psdf_samples under the build folder will activate a demo setting that reconstructs an open box soldier_fight.obj under /data folder and saves (a) the generated 3PSDF samples (the.sdffile), (b) the mesh reconstructed from the computed 3PSDF field (the.objfile), and (c) the sampling point positions (the.plyfile) to the /output folder.

To customize the usage of the program:

./gen_3psdf_samples input.obj output.sdf output_recon.obj output_sample_points.ply octree_depth [default=9] flag_writeSDF [default=1] flag_recon_3PSDF [default=1] flag_writePLY [default=1]"

(2) batch_generate: the binary that generates 3PSDF sampling points in batch.

To customize the usage of the program:

./batch_generate todo_list.txt inDir outSDFDir outObjDir outPlyDir octree_depth [default=9] \
      flag_writeSDF [default=0] flag_writeOBJ [default=1] flag_writePLY [default=1]"

TODO

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}
}