GeoGS: Geometric Prior-Guided Gaussian Splatting for robust urban reconstruction from sparse views

Gaussian Splatting (GS) has emerged as an efficient representation for large-scale urban reality capture, yet reliable reconstruction in complex built environments typically requires dense image collections with high view overlap. Under sparse observations, it often yields geometrically inaccurate and incomplete reconstructions, making external geometric priors essential for robust optimization. Geo-referenced 3D city models provide complete and metrically reliable structural priors but lack fine as-built details, whereas visual foundation models infer dense local geometric cues from images yet remain less reliable in occluded, weakly textured, or sparsely observed regions. We present GeoGS, a geometry-aware GS method that integrates worldwide-available CityGML Level of Detail (LoD) 2 models with pose-conditioned depth cues from visual foundation models for robust urban reconstruction from sparse views with known camera poses. It initializes Gaussians directly from LoD2 surfaces and optimizes them in two stages. In the Anchoring Stage, depth rendered from the 3D city model establishes a metrically consistent building backbone, while a proximity-based geometry protection mechanism prevents structural drift. In the Refinement Stage, dense visual depth is introduced with a lightweight structural anchor and adaptive weighting to recover local geometric details without compromising the established topology. Experiments on the UAV-based TUM2TWIN dataset and the street-level egenioussBench dataset demonstrate consistent improvements over state-of-the-art baselines. On TUM2TWIN, GeoGS improves average PSNR by 1.13 dB in 2D rendering assessment and reduces extracted-mesh M3C2 distance by 14.3% in 3D geometric evaluation, compared with the respective best-performing baselines. These results highlight the effectiveness of integrating structured geospatial priors with data-driven visual cues, advancing sparse-view Gaussian Splatting toward scalable and geometrically reliable urban digital twin generation. The project page is available at https://github.com/zqlin0521/GeoGS .