High-Fidelity Lightweight Mesh Reconstruction from Point Clouds (CVPR 2025 Highlight)

PyTorch implementation of the paper:

High-Fidelity Lightweight Mesh Reconstruction from Point Clouds

Chen Zhang, Wentao Wang, Ximeng Li, Xinyao Liao, Wanjuan Su, Wenbing Tao*

Setup

Python package requirements.

python == 3.9
pytorch == 1.12.1
torch_scatter
trimesh
mcubes
open3d
fpsample
tinycudann  (install via `pip install -e ./tiny-cuda-nn`)

In addition, you also need to install some C++ libraries.

Open3D == 0.16
CGAL
boost
libgmp-dev

Hash encoding dependency: The SDF/VG networks now rely on tiny-cuda-nn. Clone it next to this repo and run pip install -e ./tiny-cuda-nn (or install the published wheel) before training so that the tinycudann Python module is available.

Compiling

Before starting, you need to compile several C++ programs.

• Compile a kdtree algorithm for fast querying.

cd models/cpplib
conda activate your_env
CC=gcc CXX=gcc python setup.py build_ext --inplace

• Compile executables to construct Delaunay triangulation and generate the mesh.

cd models/delaunay_meshing

cd create_delaunay
# mkdir build
# cd build
# cmake ../
cmake ./
make

cd create_delaunay
# mkdir build
# cd build
# cmake ../
cmake ./
make

Quick Test

There is some data in the example folder for reference.
First, learn the SDF from the point cloud. example/exp/xxx/SDF is the output folder.

bash scripts/run_sdf.sh

Second, generate the mesh from point cloud and learned SDF. example/exp/xxx/VG is the output folder. The number of vertices can be specified by the parameter vertices_size in confs/vg.conf.

bash scripts/run_vg.sh

Citation

@inproceedings{zhang2025high,
  title={High-Fidelity Lightweight Mesh Reconstruction from Point Clouds},
  author={Zhang, Chen and Wang, Wentao and Li, Ximeng and Liao, Xinyao and Su, Wanjuan and Tao, Wenbing},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={11739--11748},
  year={2025}
}