Skip to content

README.md

Latest commit

 

History

History
108 lines (75 loc) · 3.87 KB

README.md

File metadata and controls

108 lines (75 loc) · 3.87 KB

[ICCV 2025] X2-Gaussian: 4D Radiative Gaussian Splatting for Continuous-time Tomographic Reconstruction

[In submission] X2-Gaussian++: Advancing 4D Radiative Gaussian Splatting for Continuous-time Tomographic Reconstruction

Project Page | Paper

We achieve genuine continuous-time CT reconstruction without phase-binning. The figure illustrates temporal variations of lung volume in 4D CT reconstructed by our X2-Gaussian.

X2-Gaussian demonstrates state-of-the-art reconstruction performance.

News

  • 2026.01.15: Training code for X2-Gaussian++ has been released. Welcome to have a try!
  • 2025.10.27: Datasets have been released here. Welcome to have a try!
  • 2025.10.17: Training code for X2-Gaussian has been released.
  • 2025.06.26: X2-Gaussian has been accepted to ICCV 2025.
  • 2025.03.27: X2-Gaussian is available on arxiv.

Installation

# Download code
git clone https://github.com/yuyouxixi/x2-gaussian.git

# Install environment
conda create -n x2_gaussian python=3.9 -y
conda activate x2_gaussian

## You can choose suitable pytorch and cuda versions here on your own.
pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
pip install -e x2_gaussian/submodules/simple-knn
## xray-gaussian-rasterization-voxelization is from https://github.com/Ruyi-Zha/r2_gaussian/tree/main/r2_gaussian/submodules/xray-gaussian-rasterization-voxelization
pip install -e x2_gaussian/submodules/xray-gaussian-rasterization-voxelization

# Install TIGRE for data generation and initialization
wget https://github.com/CERN/TIGRE/archive/refs/tags/v2.3.zip
unzip v2.3.zip
pip install TIGRE-2.3/Python --no-build-isolation

Training X2-Gaussian++

Dtaset

You can download datasets used in our paper here. We use NAF format data (*.pickle) used in SAX-NeRF.

Initialization

We have included initialization files in our dataset. You can skip this step if using our dataset.

For new data, you need to use initialize_pcd.py to generate a *.npy file which stores the point cloud for Gaussian initialization.

python initialize_pcd.py --data <path to data>

Start Training

Use train.py to train Gaussians. Make sure that the initialization file *.npy has been generated.

# Training
python train.py -s <path to data>

# Example
python train.py -s XXX/*.pickle

Training X2-Gaussian

Please refer here

Citation

If you find this work helpful, please consider citing:

@article{yu2025x,
  title={X $\^{}$\{$2$\}$ $-Gaussian: 4D Radiative Gaussian Splatting for Continuous-time Tomographic Reconstruction},
  author={Yu, Weihao and Cai, Yuanhao and Zha, Ruyi and Fan, Zhiwen and Li, Chenxin and Yuan, Yixuan},
  journal={arXiv preprint arXiv:2503.21779},
  year={2025}
}

Acknowledgement

Our code is adapted from R2-Gaussian, 4D Gaussians, X-Gaussian and TIGRE toolbox. We thank the authors for their excellent works.