| Project | Paper | arXiv | Data |

The code builds upon the fantastic 3DGS code base and borrows the data preprocessing/loading part from IDR.

We did our experiments on Ubuntu 22.04.3, CUDA 11.8, and conda environment on Python 3.7.

Clone this repository:

git clone https://github.com/turandai/gaussian_surfels.git
cd gaussian_surfels

Create conda environment:

conda env create --file environment.yml
conda activate gaussian_surfels

If you need to recompile and reinstall the CUDA rasterizer:

cd submodules/diff-gaussian-rasterization
python setup.py install && pip install .

We test our method on subsets of on DTU and BlendedMVS datasets. We select 15 scenes from DTU and 18 scenes from BlendedMVS, then preprocess and normalize the data following IDR data convention. We also adopt Omnidata to generate monocular normal prior. You can download the data from here.

To test on your own unposed data, we recommend to use COLMAP for SfM initialization. To estimate monocular normal for your own data, please follow Omnidata for additional environment setup. Download the pretrained model and run the normal estimation by:

cd submodules/omnidata
sh tools/download_surface_normal_models.sh
python estimate_normal.py --img_path path/to/your/image/directory

Note that precomputed normal of forementioned scenes from DTU and BlendedMVS are included in the downloaded dataset, so you don't have to run the normal estimation for them.

To train a scene:

python train.py -s path/to/your/data/directory

Trained model will be save in output/. To render images and reconstruct mesh from a trianed model:

python render.py -m path/to/your/trained/model --img --depth 10

We use screened Poisson surface reconstruction to extract mesh, at this line in render.py:

poisson_mesh(path, wpos, normal, color, poisson_depth, prune_thrsh)

In your output folder, xxx_plain.ply is the original mesh after Poisson reconstruction with the default depth of 10. For scenes in larger scales, you may use a higher depth level. For a smoother mesh, you may decrease the depth value. We prune the Poisson mesh with a certain threshold to remove outlying faces and output xxx_pruned.ply. This process sometimes may over-prune the mesh and results in holes. You may increase the "prune_thrsh" parameter accordingly.

To evaluate the geometry accuracy on DTU, you have to download the DTU ground truth point cloud. For BlendedMVS evaluation, we fused, denoised and normalized the ground truth multi-view depth maps to a global point cloud as the ground truth geometry, which is included in our provided dataset for download. We follow previous work to use this code to calculate the Chamfer distance between the ground truth point cloud and the reconstructed mesh.

# DTU evaluation:
python eval.py --dataset dtu --source_path path/to/your/data/directory --mesh_path path/to/your/mesh --dtu_gt_path path/to/DTU/MVSData --dtu_scanId ID
# BlendedMVS evaluation:
python eval.py --dataset bmvs --source_path path/to/your/data/directory --mesh_path path/to/your/mesh

@inproceedings{Dai2024GaussianSurfels,
  author    = {Dai, Pinxuan and Xu, Jiamin and Xie, Wenxiang and Liu, Xinguo and Wang, Huamin and Xu, Weiwei},
  title     = {High-quality Surface Reconstruction using Gaussian Surfels},
  publisher = {Association for Computing Machinery},
  booktitle = {SIGGRAPH 2024 Conference Papers},
  year      = {2024},
  doi       = {10.1145/3641519.3657441}
}