| Project Report |
Surf-LEAF (Surface-Level Extraction with Advancing Front Reconstruction Meshing) is a custom mesh extraction pipeline tailored for Gaussian Splatting reconstructions in urban environments.
Gaussian Splatting is a recent point-based scene representation that enables real-time, high-fidelity novel view synthesis across large-scale scenes. While it excels in rendering, extracting high-quality meshes from Gaussian reconstructions remains a challenge—especially in urban scenes where sharp edges, low-contrast surfaces, and sparse regions can degrade mesh fidelity.
We demonstrate improved mesh quality over existing techniques like DN-Splatter, achieving a 10% reduction in Chamfer distance on the TartanAir dataset.
Prerequisites:
- A working installation of Nerfstudio
- QED-Splatter or DN-Splatter for training Gaussian Splatting reconstructions.
- Clone the repository and its submodules by running:
git clone https://github.com/Soldann/surf-leaf.git
cd surf-leaf- Load your nerfstudio conda environment:
conda activate nerfstudio- Install the package and its dependencies:
pip install -e .The surf_leaf package provides a command line interface to run the mesh extraction pipeline. The basic usage is as follows:
surf-leaf --load_config /<Path to your run>/config.ymlThis will load the Gaussian Splatting reconstruction from the specified path and run the mesh extraction pipeline with default parameters.
A full table of parameters:
| Parameter | Type | Default | Description |
|---|---|---|---|
load_config |
Path |
REQUIRED | Path to the trained config YAML file. |
output_dir |
Path |
"./mesh_exports/" |
Path to the output directory. |
total_points |
int |
2_000_000 |
Total number of surface samples to extract. |
use_masks |
bool |
False |
If dataset has masks, use them to restrict surface sampling. |
surface_levels |
List[float] |
[0.3] |
Surface level isosurfaces to extract. |
return_normal |
"analytical", "closest_gaussian", "average" |
"closest_gaussian" |
Strategy for estimating normals from Gaussians. |
cropbox_pos |
Optional[Tuple[float, float, float]] |
None |
Position of the cropbox center (x, y, z). |
cropbox_rpy |
Optional[Tuple[float, float, float]] |
None |
Orientation of cropbox in roll, pitch, yaw (radians). |
cropbox_scale |
Optional[Tuple[float, float, float]] |
None |
Scale (size) of the cropbox. |
voxel_size |
float |
0.4 |
Voxel size used during meshing. |
margin_seam |
float |
0.04 |
Seam margin used during meshing. |
margin_discard |
float |
0.04 |
Discard margin used during meshing. |
max_edge_length |
float |
1.0 |
Maximum triangle edge length in the mesh. |
postprocess_alpha_fraction |
float |
0.001 |
Alpha wrapping ball size fraction. |
postprocess_stepsmoothnum |
int |
1 |
Number of HC Laplacian smoothing steps. |
postprocess_targetperc |
float |
0.6 |
Target reduction percentage for mesh simplification. |
For our example scene we use the TartanAir dataset, specifically the abandoned_factory/P001 scene. We use the Tartan to Nerfstudio Converter to convert the dataset to a Nerfstudio-compatible format. This includes camera poses, images, and depth images.
We use QED-Splatter to train a Gaussian Splatting reconstruction of the scene.
After training, we can extract the mesh using Surf-LEAF:
surf-leaf --load_config /outputs/qed-splatter/2025-06-23_022015/abandoned_factory/P001/config.yml \
--cropbox_scale 2.0 2.0 2.0This will extract the mesh from the Gaussian Splatting reconstruction, using a cropbox scale of 2.0 in all dimensions.
The resulting mesh will be saved in the mesh_exports directory, with the name based on the scene and timestamp.