CityGS-𝒳: A Scalable Architecture for Efficient and Geometrically Accurate Large-Scale Scene Reconstruction

arXiv 2025

Yuanyuan Gao1,2,* Hao Li1,* Jiaqi Chen1,* Zhengyu Zou1 Zhihang Zhong 2, †
Dingwen Zhang1, † Xiao Sun 2 Junwei Han1
1Northwestern Polytechnical University 2Shanghai Artificial Intelligence Laboratory

* Denotes Equal Contribution † Corresponding author

Done in Shanghai Artificial Intelligence Laboratory
arXiv Code
TL;DR: CityGS-𝒳 is a scalable 3D Gaussian Splatting framework that eliminates partition-and-merge bottlenecks through parallelized hybrid hierarchical representation (PH²-3D) and multi-GPU rendering.

We propose CityGS-𝒳, a scalable architecture for efficient and geometrically accurate large-scale scene reconstruction (left). The middle column shows qualitative results of rendered depth between our method and CityGS-v2, demonstrating CityGS-𝒳's superior geometric representation with smoother object surfaces. The top-right PSNR chart highlights CityGS-𝒳's superior reconstruction quality across various GPU configurations while significantly reducing time consumption. The bottom-right section highlights the memory efficiency of CityGS-𝒳, successfully handling high-resolution 4K rendering, while CityGS-v2 encounters out-of-memory issues.




Qualitative Comparison of Mesh and Texture

In real large-scale scene reconstruction, our method has accurate geometric details compared to the current state-of-the-art methods. Please refer to the paper for a more detailed analysis of the metrics.

Qualitative Comparison of Rendering Depths and RGBs

Our method performs well in novel view synthesis tasks while maintaining good geometry.




Quantitative Results of Novel View Synthesis

↑: higher is better, ↓: lower is better. The red, orange, and yellow colors respectively denote the best, the second best, and the third best results. Bold denotes the best result in the 'With Geometric Optimization' group. † denotes without applying the decoupled appearance encoding

Time Consumptions

We compare the time consumptions of different methods on the same dataset. The time consumptions are measured in hours.

Qualitative Comparison of Rendering Nomals

Our method demonstrates superior rendering of normals, capturing fine geometric details and smooth surfaces compared to existing techniques.

Acknowledgement

We would like to express our gratitude to the authors of the following algorithms and libraries, which have greatly inspired and supported this project:

BibTeX

If you find our work useful in your research, please consider citing: 
@article{gao2025citygsx,
        title={CityGS-$\mathcal{X}$: A Scalable Architecture for Efficient and Geometrically Accurate Large-Scale Scene Reconstruction},
        author={Yuanyuan Gao and Hao Li and Jiaqi Chen and Zhihang Zhong and Zhengyu Zou and Dingwen Zhang and Xiao Sun and Junwei Han},
        
        year={2025}
    }