ULSR-GS: Ultra Large-scale Surface Reconstruction Gaussian Splatting with Multi-View Geometric Consistency

(Retrained and re-rendered UrbanScene 3D Dataset Campus scene. Improved architectural details and road smoothing. However, the water surface remains uneven and some architectural integrity has been lost. We hope to improve this situation before the code is released.)

Abstract

While Gaussian Splatting (GS) demonstrate efficient and high-quality scene rendering and surface extraction, they fall short in handling large-scale surface extraction. To overcome this, we present ULSR-GS, a framework dedicated to high-fidelity surface extraction in ultra-large-scale scenes, addressing the limitations of existing GS-based mesh extraction methods. Specifically, we propose a point-to-photo partitioning approach combined with a multi-view optimal view matching principle to select the best training images for each sub-region. Additionally, during training, ULSR-GS employs a densification strategy based on multi-view geometric consistency to enhance rendering and surface extraction details. Experimental results demonstrate that ULSR-GS outperforms other state-of-the-art GS-based works on large-scale benchmark datasets, significantly improving both rendering quality and surface extraction accuracy in complex urban environments.

Partition Example

Merge Partition Results

Virtual Larg-Scale Scene (Matrix City Small City)

Comparison With GS-based SOTA

Please click the videos for better view.

Ours

2DGS

Ours

PGSR

Ours

GOF

Ours

SuGaR

Images

BibTeX

@misc{li2024ulsrgsultralargescalesurface,
      title={ULSR-GS: Ultra Large-scale Surface Reconstruction Gaussian Splatting with Multi-View Geometric Consistency}, 
      author={Zhuoxiao Li and Shanliang Yao and Qizhong Gao and Angel F. Garcia-Fernandez and Yong Yue and Xiaohui Zhu},
      year={2024},
      eprint={2412.01402},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.01402}, 
}