2D Gaussian Splatting for Geometrically Accurate Radiance Fields

Project page | Paper | Video | Surfel Rasterizer (CUDA) | Surfel Rasterizer (Python) | DTU+COLMAP (3.5GB) | SIBR Viewer Pre-built for Windows | Web Viewer

This repo contains the official implementation for the paper "2D Gaussian Splatting for Geometrically Accurate Radiance Fields". Our work represents a scene with a set of 2D oriented disks (surface elements) and rasterizes the surfels with perspective correct differentiable raseterization. Our work also develops regularizations that enhance the reconstruction quality. We also devise meshing approaches for Gaussian splatting.

⭐ New Features

• 2024/07/20: Web-based viewer GaussianSplats3D also supports 2DGS. Thanks to Mark Kellogg
• 2024/07/19: Colab Notebook is supported! Thanks to atakan-topaloglu
• 2024/06/10: SIBR Viewer is supported!
• 2024/06/05: Remote Viewer based on Viser is supported! Thanks to HwanHeo.
• 2024/05/30: Fixed a bug related to unbounded meshing. The foreground mesh quality should now be consistent with the bounded mesh.
• 2024/05/17: Improve training speed by 30%~40% through the cuda operator fusing. Please update the diff-surfel-rasterization submodule if you have already installed it.

  git submodule update --remote  
  pip install submodules/diff-surfel-rasterization

• 2024/05/05: Important updates - Now our algorithm supports unbounded mesh extraction! Our key idea is to contract the space into a sphere and then perform adaptive TSDF truncation.

SIBR Viewer

2DGS.Viewer.mp4

The Pre-built Viewer for Windows can be found here. If you use Ubuntu or want to check the viewer usage, please refer to GS Monitor.

How to use

Firstly open the viewer,

<path to downloaded/compiled viewer>/bin/SIBR_remoteGaussian_app_rwdi

and then

# Monitor the training process
python train.py -s <path to COLMAP or NeRF Synthetic dataset> 
# View the trained model
python view.py -s <path to COLMAP or NeRF Synthetic dataset> -m <path to trained model> 

Installation

# download
git clone https://github.com/hbb1/2d-gaussian-splatting.git --recursive

# if you have an environment used for 3dgs, use it
# if not, create a new environment
conda env create --file environment.yml
conda activate surfel_splatting

Training

To train a scene, simply use

python train.py -s <path to COLMAP or NeRF Synthetic dataset>

Commandline arguments for regularizations

--lambda_normal  # hyperparameter for normal consistency
--lambda_distortion # hyperparameter for depth distortion
--depth_ratio # 0 for mean depth and 1 for median depth, 0 works for most cases

Tips for adjusting the parameters on your own dataset:

• For unbounded/large scenes, we suggest using mean depth, i.e., depth_ratio=0, for less "disk-aliasing" artifacts.

Testing

Bounded Mesh Extraction

To export a mesh within a bounded volume, simply use

python render.py -m <path to pre-trained model> -s <path to COLMAP dataset> 

Commandline arguments you should adjust accordingly for meshing for bounded TSDF fusion, use

--depth_ratio # 0 for mean depth and 1 for median depth
--voxel_size # voxel size
--depth_trunc # depth truncation

If these arguments are not specified, the script will automatically estimate them using the camera information.

Unbounded Mesh Extraction

To export a mesh with an arbitrary size, we devised an unbounded TSDF fusion with space contraction and adaptive truncation.

python render.py -m <path to pre-trained model> -s <path to COLMAP dataset> --mesh_res 1024

Quick Examples

Assuming you have downloaded MipNeRF360, simply use

python train.py -s <path to m360>/<garden> -m output/m360/garden
# use our unbounded mesh extraction!!
python render.py -s <path to m360>/<garden> -m output/m360/garden --unbounded --skip_test --skip_train --mesh_res 1024
# or use the bounded mesh extraction if you focus on foreground
python render.py -s <path to m360>/<garden> -m output/m360/garden --skip_test --skip_train --mesh_res 1024

If you have downloaded the DTU dataset, you can use

python train.py -s <path to dtu>/<scan105> -m output/date/scan105 -r 2 --depth_ratio 1
python render.py -r 2 --depth_ratio 1 --skip_test --skip_train

Custom Dataset: We use the same COLMAP loader as 3DGS, you can prepare your data following here.

Full evaluation

We provide scripts to evaluate our method of novel view synthesis and geometric reconstruction.

Explanation of Performance Differences to the Paper

We have re-implemented the repository for improved efficiency, which has slightly impacted performance compared to the original paper. Two factors have influenced this change:

• 📈 We fixed some minor bugs, such as a half-pixel shift in TSDF fusion, resulting in improved geometry reconstruction.

• 📉 We removed the gradient of the low-pass filter used for densification, which reduces the number of Gaussians. As a result, the PSNR has slightly dropped, but we believe this trade-off is worthwhile for real-world applications.

You can report either the numbers from the paper or from this implementation, as long as they are discussed in a comparable setting.

Novel View Synthesis

For novel view synthesis on MipNeRF360 (which also works for other colmap datasets), use

python scripts/mipnerf_eval.py -m60 <path to the MipNeRF360 dataset>

We provide Evaluation Results (Pretrained, Images).

Table Results

Geometry reconstruction

For geometry reconstruction on DTU dataset, please download the preprocessed data from Drive or Hugging Face. You also need to download the ground truth DTU point cloud.

python scripts/dtu_eval.py --dtu <path to the preprocessed DTU dataset>   \
     --DTU_Official <path to the official DTU dataset>

We provide Evaluation Results (Pretrained, Meshes).

Table Results

Chamfer distance on DTU dataset (lower is better)

	24	37	40	55	63	65	69	83	97	105	106	110	114	118	122	Mean
Paper	0.48	0.91	0.39	0.39	1.01	0.83	0.81	1.36	1.27	0.76	0.70	1.40	0.40	0.76	0.52	0.80
Reproduce	0.46	0.80	0.33	0.37	0.95	0.86	0.80	1.25	1.24	0.67	0.67	1.24	0.39	0.64	0.47	0.74

For geometry reconstruction on TnT dataset, please download the preprocessed TnT_data. You also need to download the ground truth TnT_GT, including ground truth point cloud, alignments and cropfiles.

Due to historical issue, you should use open3d==0.10.0 for evaluating TNT.

# use open3d 0.18.0, skip metrics
python scripts/tnt_eval.py --TNT_data <path to the preprocessed TNT dataset>   \
     --TNT_GT <path to the official TNT evaluation dataset> --skip_metrics

# use open3d 0.10.0, skip traing and rendering
python scripts/tnt_eval.py --TNT_data <path to the preprocessed TNT dataset>   \
     --TNT_GT <path to the official TNT evaluation dataset> --skip_training --skip_rendering

We provide Evaluation Results (Pretrained, Meshes).

Table Results

F1 scores on TnT dataset (higher is better)

	Barn	Caterpillar	Ignatius	Truck	Meetingroom	Courthouse	Mean
Reproduce	0.41	0.23	0.51	0.45	0.17	0.15	0.32

FAQ

• Training does not converge. If your camera's principal point does not lie at the image center, you may experience convergence issues. Our code only supports the ideal pinhole camera format, so you may need to make some modifications. Please follow the instructions provided here to make the necessary changes. We have also modified the rasterizer in the latest commit to support data accepted by 3DGS. To avoid further issues, please update to the latest commit.

• No mesh / Broken mesh. When using the Bounded mesh extraction mode, it is necessary to adjust the depth_trunc parameter to perform TSDF fusion to extract meshes. On the other hand, Unbounded mesh extraction does not require tuning the parameters but is less efficient.

• Can 3DGS's viewer be used to visualize 2DGS? Technically, you can export 2DGS to 3DGS's ply file by appending an additional zero scale. However, due to the inaccurate affine projection of 3DGS's viewer, you may see some distorted artefacts. We are currently working on a viewer for 2DGS, so stay tuned for updates.

Acknowledgements

This project is built upon 3DGS. The TSDF fusion for extracting mesh is based on Open3D. The rendering script for MipNeRF360 is adopted from Multinerf, while the evaluation scripts for DTU and Tanks and Temples dataset are taken from DTUeval-python and TanksAndTemples, respectively. The fusing operation for accelerating the renderer is inspired by Han's repodcue. We thank all the authors for their great repos.

Citation

If you find our code or paper helps, please consider citing:

@inproceedings{Huang2DGS2024,
    title={2D Gaussian Splatting for Geometrically Accurate Radiance Fields},
    author={Huang, Binbin and Yu, Zehao and Chen, Anpei and Geiger, Andreas and Gao, Shenghua},
    publisher = {Association for Computing Machinery},
    booktitle = {SIGGRAPH 2024 Conference Papers},
    year      = {2024},
    doi       = {10.1145/3641519.3657428}
}