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

Project page

1. We change to calculate depth using per-pixel cosine values (Eq. 14: $d=cos\theta\ t^*$). Additionally, an option is provided to directly render the camera space coordinate map. We utilize the inverse of affine approximation to transform intersections from ray space to camera space, and these transformed points are then used in computing normal consistency loss.
2. The depth distortion loss has been eliminated from our training process. Currently, we rely solely on normal consistency loss for geometry regularization. We believe future techniques will enhance performance even further.

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

1. create an environment

conda create -n radegs python=3.9
conda activate radegs

2. install pytorch and other dependencies.

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

3. install submodules

pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn/

# tetra-nerf for Marching Tetrahedra
cd submodules/tetra-triangulation
conda install cmake
conda install conda-forge::gmp
conda install conda-forge::cgal
cmake .
# you can specify your own cuda path
# export CPATH=/usr/local/cuda-11.3/targets/x86_64-linux/include:$CPATH
make 
pip install -e .

We use preprocessed DTU dataset from 2DGS for training. And we follow GOF to evaluate the geometry. Point clouds from the DTU dataset need saved to dtu_eval/Offical_DTU_Dataset for the geometry evaluation. We use preprocessed Tanks and Temples dataset from GOF. For evalution, please download ground truth point cloud, camera poses, alignments and cropfiles from Tanks and Temples dataset. The ground truth dataset should be organized as:

GT_TNT_dataset
│
└─── Barn
│   │
|   └─── Barn.json
│   │
|   └─── Barn.ply
│   │
|   └─── Barn_COLMAP_SfM.log
│   │
|   └─── Barn_trans.txt
│ 
└─── Caterpillar
│   │
......

# training
python train.py -s <path to DTU dataset> -m <output folder> -r 2 --use_decoupled_appearance
# mesh extraction
python mesh_extract.py -s <path to DTU dataset> -m <output folder> -r 2
# evaluation
python evaluate_dtu_mesh.py -s <path to DTU dataset> -m <output folder>

# training
python train.py -s <path to preprocessed TNT dataset> -m <output folder> -r 2 --eval --use_decoupled_appearance
# mesh extraction
python mesh_extract_tetrahedra.py -s <path to preprocessed TNT dataset> -m <output folder> -r 2 --eval
# evaluation
python eval_tnt/run.py --dataset-dir <path to GT TNT dataset> --traj-path <path to preprocessed TNT COLMAP_SfM.log file> --ply-path <output folder>/recon.ply

python train.py -s <path to COLMAP or NeRF Synthetic dataset> --eval
python render.py -m <path to pre-trained model> -s <path to dataset>
python metrics.py -m <path to trained model> # Compute error metrics on renderings

Our model can directly render coordinate map for training, without the need to first render a depth map and then convert it. This feature can be activated by including --use_coord_map in the argument list of 'train.py'.

Current viewer in this repository is very similar to the original Gaussian Splatting viewer (with small modifications for 3D filters). You can build and use it in the same way as Gaussian Splatting.

We build this project based on Gaussian Splatting.

We incorporate the filters proposed in Mip-Splatting.

We incorporate the loss functions of 2D GS and use the preprocessed DTU dataset.

We incorporate the densification strategy, evalution and decoupled appearance modeling form GOF and use the preprocessed TNT dataset.

The evaluation scripts for the DTU and Tanks and Temples datasets are sourced from DTUeval-python and TanksAndTemples, respectively.

We thank the authors of Gaussian Splatting, Mip-Splatting, 2D GS, GOF， and the repos for their great works.