Awesome

<p align="center" /> <h1 align="center">Binocular-Guided 3D Gaussian Splatting with ViewConsistency for Sparse View Synthesis </h1> <p align="center"> <strong>Liang Han</strong> · <a href="https://junshengzhou.github.io/"><strong>Junsheng Zhou</strong></a> · <a href="https://yushen-liu.github.io/"><strong>Yu-Shen Liu</strong></a> · <a href="https://h312h.github.io/"><strong>Zhizhong Han</strong></a> </p> <h2 align="center">NeurIPS 2024</h2> <h3 align="center"><a href="https://arxiv.org/abs/2410.18822">Paper</a> | <a href="https://hanl2010.github.io/Binocular3DGS/">Project Page</a></h3> <div align="center"></div> <p align="center"> <img src="assets/pipeline.png" width="780" /> </p>

We leverage dense initialization for achieving Gaussian locations, and optimize the locations and Gaussian attributes with three constraints or strategies:

<ul> <li> Binocular Stereo Consistency Loss. We construct a binocular view pair by translating an input view with camera positions, where we constrain on the view consistency of binocular view pairs in a self-supervised manner.</li> <li> Opacity Penalty Strategy is designed to decay the Gaussian opacity during training for regularizing them. </li> <li> The commonly-used Color Reconstruction Loss. </li> </ul>

Installation

Clone Binocular3DGS

git clone --recursive https://github.com/hanl2010/Binocular3DGS.git

Setup Anaconda Environment

conda create -n binocular3dgs python=3.10
conda activate binocular3dgs
pip install -r requirements.txt
pip install submodules/diff-gaussian-rasterization
pip install submodules/simple-knn

Dataset

Download the processed datasets: LLFF and DTU
Download the NeRF Synthetic dataset from here

Checkpoints

We use the pre-trained PDCNet+ to generate dense initialization point clouds. The pre-trained PDCNet+ model can be downloaded here.

Put the pre-trained model in submodules/dense_matcher/pre_trained_models

Training and Evaluation

LLFF dataset

python script/run_llff.py

DTU dataset

python script/run_dtu.py

NeRF Synthetic dataset (Blender)

python script/run_blender.py

Citation

If you find our code or paper useful, please consider citing

@inproceedings{han2024binocular,
    title = {Binocular-Guided 3D Gaussian Splatting with View Consistency for Sparse View Synthesis},
    author = {Han, Liang and Zhou, Junsheng and Liu, Yu-Shen and Han, Zhizhong},
    booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
    year = {2024}
}

Acknowledgement

This project is built upon gaussian-splatting with simple-knn and a modified diff-gaussian-rasterization. The scripts for generating videos are borrowed from DNGaussian and the scripts for dense matching are from PDCNet+. Thanks for these great projects.