SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction

NeurIPS 2025

Paper | Pretrained Models

Installation

Create a Python 3.10 environment and install the dependencies:

git clone https://github.com/Simon-Dcs/Surfel_Splat.git
cd Surfel_Splat
conda create -n surfelsplat python=3.10
conda activate surfelsplat
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118
pip install --no-build-isolation -r requirements_w_versions.txt

• If you encounter problems like ModuleNotFoundError: No module named 'pkg_resources',try to run pip install "setuptools<82"
• If you encounter numpy version problems, please run pip install numpy==1.26.3

Checkpoints

Download the SurfelSplat checkpoint here and place it under checkpoints/.

The UniMatch backbone weights are also required:

mkdir -p checkpoints
wget 'https://s3.eu-central-1.amazonaws.com/avg-projects/unimatch/pretrained/gmdepth-scale1-resumeflowthings-scannet-5d9d7964.pth' -P checkpoints

Datasets

The codebase uses chunked dataset files in the same general style as pixelSplat. You should prepare your datasets into the format expected by the loaders in src/dataset/.

We present our preprocessed dataset here. By default, the config points to:

datasets/torch_data

Update the dataset roots in the experiment config if your local dataset layout differs:

• config/experiment/re10k.yaml

Inference

Scene generation

The main entrypoint is:

• generate.sh

Available handles:

• SCENE_ID: integer scene id, for example 24
• CONTEXT_VIEWS: comma-separated context view ids
• TARGET_VIEWS: comma-separated target view ids
• CHECKPOINT_PATH: checkpoint path
• CUDA_DEVICE: GPU id

Example:

SCENE_ID=24 \
CONTEXT_VIEWS=0,2,9 \
TARGET_VIEWS=4,2 \
CHECKPOINT_PATH=checkpoints/checkpoint.ckpt \
CUDA_DEVICE=0 \
bash generate.sh

You can also refer to dataset-specific script such as generate_dtu.sh and generate_blendmvs.sh.

Training

The lightweight training wrapper is:

• train.sh

Example:

CUDA_DEVICES=0 \
BATCH_SIZE=1 \
bash train.sh

Before training, comment out the code at line 276 in src/dataset/dataset_re10k.py and line 529 in src/model/encoder/encoder_costvolume.py, since this release is configured to run with a global scale factor of 1/200. Adjust the experiment config, dataset roots, and batch size according to your hardware and dataset setup.

Mesh Reconstruction

The mesh reconstruction utility is:

• src/mesh/gs2mesh.py

Example:

GS2MESH_INPUT_PLY=point_clouds/data/000000_scan24_train/gaussians.ply \
GS2MESH_OUTPUT_MESH=point_clouds/data/000000_scan24_train/scene24_mesh.ply \
GS2MESH_TEMP_MESH=point_clouds/data/000000_scan24_train/temp.ply \
GS2MESH_DOUBLE_SIDED=true \
GS2MESH_VISUALIZE=true \
python src/mesh/gs2mesh.py

Evaluation

To evaluate a reconstructed DTU mesh, run:

python src/mesh/evaluate_single_scene.py \
  --input_mesh /path/to/mesh \
  --scan_id your_id \
  --output_dir /path/to/output \
  --DTU /path/to/DTU

The official DTU ground-truth data can be found here.

BibTeX

@inproceedings{daisurfelsplat,
  title={SurfelSplat: Learning Efficient and Generalizable Gaussian Surfel Representations for Sparse-View Surface Reconstruction},
  author={Dai, Chensheng and Zhang, Shengjun and Chen, Min and Duan, Yueqi},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems}
}