ULSR-GS: Urban large-scale surface reconstruction Gaussian Splatting with multi-view geometric consistency

Zhuoxiao Li, Shanliang Yao, Rongjun Qin, Wufan Zhao, Jason Ralph, Xiaohui Zhu

Abstract: Recent advances in 2D Gaussian Splatting (2DGS) have demonstrated compelling rendering efficiency and mesh extraction capabilities. However, its application to large-scale aerial photogrammetry, especially using oblique UAV imagery, remains limited due to three primary challenges: (1) suboptimal image selection in scene partitioning strategies failing to scale effectively; (2) densification pipelines that rely primarily on single-view constraints, resulting in under-reconstructions and loss of fine geometric detail; and (3) the absence of multi-view geometric consistency constraints leading to surface artifacts and inconsistencies. To address these limitations, we propose ULSR-GS, a novel method tailored for high-resolution surface reconstruction in urban-scale environments. Firstly, we propose a point-to-photo partitioning strategy that segments the scene based on the sparse SfM point cloud and assigns only the most relevant images to each sub-region, which resolves key scalability bottlenecks. Secondly, we propose a multi-view guided densification strategy that enforces adaptive geometric consistency across views, overcoming the limitations of single-view-based densifications. Lastly, we introduce consistency-aware loss functions that explicitly regulate depth and normal alignment across views, significantly enhancing surface fidelity. Extensive experiments on large-scale aerial benchmark datasets demonstrate that ULSR-GS consistently outperforms existing single- and multi-GPU Gaussian Splatting methods. Furthermore, compared to MVS pipelines, our approach achieves comparable or superior geometric quality while being substantially more time-efficient, making it a practical solution for scalable 3D modeling in digital twin and urban mapping applications. Project page: https://ulsrgs.github.io.