RaDe-GS: Rasterizing Depth in Gaussian Splatting

Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, Ping Tan

Project page

News!

• The paper has been accepted for publication in ACM Transactions on Graphics (TOG)!
• We incorporate the multi-view regularization from PGSR.

1. Installation

Clone this repository.

git clone https://github.com/HKUST-SAIL/RaDe-GS.git --recursive

Create an environment

conda create -n radegs python=3.12
conda activate radegs

Install pytorch and other dependencies.

pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt

Install submodules

pip install submodules/diff-gaussian-rasterization --no-build-isolation
pip install submodules/warp-patch-ncc --no-build-isolation
pip install submodules/simple-knn/ --no-build-isolation
pip install git+https://github.com/rahul-goel/fused-ssim/ --no-build-isolation

# tetra-nerf for Marching Tetrahedra
conda install conda-forge::cgal
pip install submodules/tetra_triangulation/ --no-build-isolation

---

2. Data Preparation

DTU

We train on the preprocessed DTU dataset from 2DGS:
https://surfsplatting.github.io/

For geometry evaluation, download the official DTU point clouds and place them under:

dtu_eval/Offical_DTU_Dataset

DTU dataset page: https://roboimagedata.compute.dtu.dk/?page_id=36

Tanks and Temples (TnT)

Please follow PGSR to preprocess the TnT dataset. For evaluation, download the GT point clouds, camera poses, alignments, and crop files from:
https://www.tanksandtemples.org/download/

Expected structure:

GT_TNT_dataset/
  Barn/
    images/
      000001.jpg
      000002.jpg
      ...
    sparse/
      0/
        ...
    Barn.json
    Barn.ply
    Barn_COLMAP_SfM.log
    Barn_trans.txt
  Caterpillar/
    ...

Objaverse

For depth and normal evaluation, we render multi-view images from Objaverse assets and export the corresponding ground-truth depth maps and surface normal maps. The rendered dataset can be downloaded from this link.

---

3. Training & Evaluation

Below are example commands for training, mesh extraction, rendering, and evaluation.

DTU

# Training
python train.py -s <path_to_dtu> -m <output_dir> -r 2 --use_decoupled_appearance 3

# Mesh extraction
python mesh_extract.py -m <output_dir>

# Evaluation
python evaluate_dtu_mesh.py -m <output_dir>

Tanks and Temples (TnT)

# Training
python train.py -s <path_to_preprocessed_tnt> -m <output_dir> -r 2 --use_decoupled_appearance 3

# Mesh extraction
python mesh_extract_tnt.py -m <output_dir>

# Evaluation
python eval_tnt/run.py \
  --dataset-dir <path_to_gt_tnt> \
  --traj-path <path_to_COLMAP_SfM.log> \
  --ply-path <output_dir>/recon_post.ply \
  --out-dir <output_dir>/mesh

Novel View Synthesis

# Training
python train.py -s <path_to_dataset> -m <output_dir> --eval

# Rendering
python render.py -m <output_dir>

# Evaluation
python metrics.py -m <output_dir>

Objaverse

# Training
python train.py -s <path_to_dataset> -m <output_dir> --eval

# Evaluation
python geometry_metric.py -m <output_dir>

---

4. Viewer

Current viewer in this repository is very similar to the original Gaussian Splatting viewer, with minor updates for newer library versions and for loading 3D Gaussian models. You can build and use it in the same way as Gaussian Splatting.

5. Acknowledgements

This project is built upon the original implementation of 3D Gaussian Splatting (3DGS): https://github.com/graphdeco-inria/gaussian-splatting.

We integrate components and ideas from several recent works, including the filtering strategy from Mip-Splatting, and regularization terms from 2DGS and PGSR.

We also incorporate the densification strategy proposed in GOF, and adopt decoupled appearance modeling practices inspired by 3DGS, GOF, and PGSR.

For geometric evaluation, we use the DTU and Tanks and Temples evaluation toolboxes from DTUeval-python (https://github.com/jzhangbs/DTUeval-python) and the TanksAndTemples Python evaluation scripts (https://github.com/isl-org/TanksAndTemples/tree/master/python_toolbox/evaluation), respectively.

We thank the authors of these projects for making their code publicly available.