Takeshi Noda · Yu-Shen Liu · Zhizhong Han
Overview of our method. Given 3D Gaussians 𝑔, we employ a distance field specified by a fused TSDF grid as our prior 𝑓ᵗ. With 𝑓ᵗ, we define a bandwidth around the surface and iteratively refine 𝑓ᵗ with updated depth renderings. We also apply Gaussian geometric constraints (GC) that are related to the interpolated distance 𝑠, Gaussian centers μ, and TSDF gradients ∇𝑓ᵗ for high-fidelity surface reconstruction.
- Clone GSPrior
git clone --recursive https://github.com/takeshie/GSPrior.git
- Setup Anaconda Environment
conda create -n gsprior python=3.8
conda activate gsprior
pip install -r requirements.txt
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn
- Download the NeRF-Synthetic dataset from (https://drive.google.com/drive/folders/1cK3UDIJqKAAm7zyrxRYVFJ0BRMgrwhh4)
- Download the DTU dataset datasets from (https://roboimagedata.compute.dtu.dk/?page_id=36)
- Download the TNT dataset datasets from (https://www.tanksandtemples.org/download/)
- Download the MIP-NeRF 360 datasets from (https://jonbarron.info/mipnerf360/)
We provide a simple example for training and evaluation on the DTU dataset.
python train.py -s <path_to_scene> -m <path_to_output>After training, use render.py for rendering and evaluation:
python render.py -s <dataset_path>/<scene_name> -m <output_path>/<scene_name>For training and evaluating multiple scenes, we follow the evaluation protocol of PGSR and provide the corresponding scripts in the scripts directory.
bash scripts/<script_name>.shPlease modify the dataset paths, output paths, scene IDs, and GPU settings in the scripts according to your local environment. After training and evaluation, the final results include the visualization results of the TSDF and TSDF prior. We provide reference outputs in the out directory. The generated results may include rendered images, evaluation outputs, TSDF visualization results, and TSDF-prior visualization results.
If you find our code or paper useful, please consider citing
@article{noda20263d,
title={3D Gaussian Splatting with Self-Constrained Priors for High Fidelity Surface Reconstruction},
author={Noda, Takeshi and Liu, Yu-Shen and Han, Zhizhong},
journal={arXiv preprint arXiv:2603.19682},
year={2026}
}This project is built upon PGSR and GS-Pull. Thanks for these great projects.