UrbanGS: A Scalable and Efficient Architecture for Geometrically Accurate Large-Scene Reconstruction

TL;DR

UrbanGS addresses the challenge of large-scale 3D reconstruction with 3D Gaussian Splatting by introducing a depth-consistent D-Normal regularizer that unifies depth-derived normals with external pseudo-depth priors to update all Gaussian parameters. It further introduces Spatially Adaptive Gaussian Pruning to remove redundant primitives and a unified partitioning strategy to enable boundary-artifact-free, parallelizable reconstruction. On Mill-19, UrbanScene3D, and GauU-Scene, the method achieves state-of-the-art rendering quality and geometric accuracy while improving training efficiency and memory usage. Together, these components provide a scalable, high-fidelity solution for urban-scale 3D reconstruction with practical implications for real-time visualization and mapping.

Abstract

While 3D Gaussian Splatting (3DGS) enables high-quality, real-time rendering for bounded scenes, its extension to large-scale urban environments gives rise to critical challenges in terms of geometric consistency, memory efficiency, and computational scalability. To address these issues, we present UrbanGS, a scalable reconstruction framework that effectively tackles these challenges for city-scale applications. First, we propose a Depth-Consistent D-Normal Regularization module. Unlike existing approaches that rely solely on monocular normal estimators, which can effectively update rotation parameters yet struggle to update position parameters, our method integrates D-Normal constraints with external depth supervision. This allows for comprehensive updates of all geometric parameters. By further incorporating an adaptive confidence weighting mechanism based on gradient consistency and inverse depth deviation, our approach significantly enhances multi-view depth alignment and geometric coherence, which effectively resolves the issue of geometric accuracy in complex large-scale scenes. To improve scalability, we introduce a Spatially Adaptive Gaussian Pruning (SAGP) strategy, which dynamically adjusts Gaussian density based on local geometric complexity and visibility to reduce redundancy. Additionally, a unified partitioning and view assignment scheme is designed to eliminate boundary artifacts and optimize computational load. Extensive experiments on multiple urban datasets demonstrate that UrbanGS achieves superior performance in rendering quality, geometric accuracy, and memory efficiency, providing a systematic solution for high-fidelity large-scale scene reconstruction.

Figures (12)

• Figure 1: We propose UrbanGS, a scalable framework for high-fidelity large-scale scene reconstruction. Left: It reconstructs complex urban environments from multi-view RGB images, capturing fine details like trees, buildings, and roads. Middle: Compared with CityGS-v2 citygsv2 and VCR-Gaus VCR, by comparing rendered depth maps, our method can intuitively demonstrate its geometric advantages in terms of the surface smoothness of objects. Top-right: Our Spatially Adaptive Gaussian Pruning enables significant model compression while preserving quality. Bottom-right: UrbanGS efficiently reconstructs large scenes on A5000 GPUs, whereas VCR-Gaus VCR fails due to out-of-memory issues.
• Figure 2: UrbanGS training pipeline and core components. (a) Training Pipeline: Starting from coarse global Gaussians, we apply spatially adaptive Gaussian pruning to obtain compact priors, contract and partition the scene into blocks, assign camera views using geometric and SSIM-based criteria, and refine all blocks in parallel before merging them into a unified large-scale 3D Gaussian scene. (b) Depth-Consistent D-Normal Regularization: 3D Gaussians are rendered to depth and normal maps, depth is converted to D-normals and jointly supervised with pseudo-depth and pseudo-normal priors from pretrained models via the loss $\mathcal{L}_n + \mathcal{L}_{dn} + w_d \mathcal{L}_{id}$, yielding stable and globally consistent geometry. (c) Spatially Adaptive Gaussian Pruning: Global Gaussians are discretized into a voxel grid, where per-cell importance $\omega_{v,i}$ and view-dependent cues are fused into pruning scores to remove redundant Gaussians and obtain an efficient yet accurate representation.
• Figure 3: Qualitative results of ours and other methods in image rendering on Mill-19 monosdf and Urbanscene3D urbanscene.
• Figure 4: Qualitative mesh and texture comparison between SOTA and our method on GauU-Scene dataset GauU-scene.
• Figure 5: Experimental results on the Rubble dataset monosdf demonstrate that the proposed method outperforms comparative approaches in terms of PSNR while achieving superior training efficiency.