NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting

Yulong Zheng¹ Zicheng Jiang¹ Shengfeng He² Yandu Sun¹ Junyu Dong¹ Huaidong Zhang³ Yong Du^{1, *}

¹Ocean University of Chinay ²Singapore Management University ³South China University of Technology

^*corresponding author

Environmental Setups

Tested on Ubuntu 18.04, CUDA 11.8, PyTorch 2.0.0

conda env create -n nexus python=3.10
conda activate nexus

Install Pytorch

pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118

Install submodules

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

Running

Taking LLFF as an example, the dataset folder structure is as follows:

├── datasets
    ├── LLFF
        ├── scene
            ├── sparse
            ├── images 
            ├── images_8
            ├── 3_views
                ├── flow

LLFF

Download LLFF dataset: Link.

Download the LLFF optical flow processed by FlowFormer++ from the Link.

Run using the following script:

sh scripts/run_llff.sh 0

DTU

TODO

MipNeRF-360

TODO

Citation

If you find our work useful for your project, please consider citing the following paper.

@article{zheng2025nexusgs,
  title={NexusGS: Sparse View Synthesis with Epipolar Depth Priors in 3D Gaussian Splatting},
  author={Zheng, Yulong and Jiang, Zicheng and He, Shengfeng and Sun, Yandu and Dong, Junyu and Zhang, Huaidong and Du, Yong},
  journal={arXiv preprint arXiv:2503.18794},
  year={2025}
}

Acknowledgement

Special thanks to the following awesome projects!