SDFoam: Signed-Distance Foam for Explicit Surface Reconstruction

Antonella Rech, Nicola Conci, Nicola Garau

SDFoam jointly optimizes an implicit Signed Distance Field (SDF) and an explicit Voronoi Diagram, or foam, for ray-traced scene representation and mesh reconstruction. The model keeps the speed and explicit structure of RadiantFoam while using SDF supervision and regularization to improve surface quality, reduce floaters, and extract cleaner meshes.

This branch contains the current setup used for training, live viewing, and SDF/alpha-based Voronoi mesh extraction.

Overview

NeRF-style methods produce strong novel-view synthesis but often make mesh extraction difficult. RadiantFoam organizes radiance with explicit Voronoi cells and ray tracing, but precise surface reconstruction can still suffer from holes and floaters. SDFoam addresses this by learning an SDF together with the Voronoi foam. The SDF gives a metric surface signal, and the foam provides an explicit non-overlapping structure for rendering and mesh extraction.

Requirements

The current workflow has been tested on Windows with:

• Windows 10/11 (though easily portable to Linux)
• NVIDIA GPU with CUDA support
• CUDA Toolkit 13.2, with nvcc available on PATH
• Visual Studio 2022 Build Tools with the C++ desktop workload
• Python 3.12
• uv
• Git

The setup installs PyTorch CUDA 13.2 wheels:

• torch==2.12.0+cu132
• torchvision==0.27.0+cu132

Quick Setup

From a fresh checkout:

git clone --recursive https://github.com/mmlab-cv/SDFoam.git
cd SDFoam

uv venv .venv
.venv\Scripts\activate

uv pip install -r requirements.txt
uv pip install --reinstall torch torchvision --index-url https://download.pytorch.org/whl/cu132
uv pip install .

python train.py -c configs\dtu_scan.yaml --viewer

Training

Training uses a YAML config:

python train.py -c configs\dtu_scan.yaml

To launch the native training viewer at the same time:

python train.py -c configs\dtu_scan.yaml --viewer

Runs are saved under output/<scene_name>_<mode>@<timestamp>/. A completed or intermediate run should contain at least:

output/<run>/config.yaml
output/<run>/model.pt

Use the generated config.yaml when opening a trained run in a viewer.

SDF + Alpha Voronoi Viewer

This is the current mesh-inspection viewer. It filters Voronoi sites using SDF and alpha values, shows the selected geometry, and exports colored mesh files.

python gui.py -c .\output\<run>\config.yaml

Headless Mesh Extraction

If you want to extract the Voronoi mesh without opening the Open3D GUI, use extract_mesh.py with the checkpoint config:

python extract_mesh.py -c .\output\<run>\config.yaml

The default extraction keeps seeds in this range:

-0.02 < SDF < 0.05
0.096 < alpha < 1.0

You can override the thresholds and output name:

python extract_mesh.py -c .\output\<run>\config.yaml --sdf_min -0.02 --sdf_max 0.05 --alpha_min 0.096 --alpha_max 1.0 --out_base scan65_mesh

The script writes colored mesh files using the selected base name:

scan65_mesh.ply
scan65_mesh.obj
scan65_mesh.mtl

CMake Development Build

Use this path if you are editing C++ or CUDA code and want incremental rebuilds:

mkdir build
cd build
cmake ..
cmake --build . --config Release --target install
cd ..
python train.py -c configs\dtu_scan.yaml --viewer

uv pip install . and the direct CMake flow now use the same CMake project and install the Python bindings into the local package.

Credits

This repository structure is based on the Radiant Foam codebase: theialab/radfoam.

BibTeX

@inproceedings{rech2026sdfoam,
  title={SDFoam: Signed-Distance Foam for explicit surface reconstruction},
  author={Rech, Antonella and Conci, Nicola and Garau, Nicola},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={300--308},
  year={2026}
}