City-Mesh3R-v2: Efficient and High-Quality City-Scale 3D Mesh Reconstruction

High-fidelity, simulation-ready city-scale 3D mesh reconstruction from multi-view RGB images remains an open challenge due to the scale and geometric complexity of recovering fine-grained urban surfaces solely from 2D image cues. Existing scalable approaches typically adopt hierarchical pipelines based on scene partitioning, per-partition Gaussian or mesh reconstruction, and subsequent global geometry merging. However, these methods often produce irregular mesh surfaces and structural distortions that limit their suitability for downstream simulation applications. In this paper, we propose a novel mesh optimization framework based on a three-stage hybrid optimization strategy that jointly improves convergence, geometric fidelity, and surface regularity while preserving fine-grained urban details. We further introduce a detail-aware surface regularization loss that suppresses local mesh artifacts, including noise and geometric distortions. City-Mesh3R-v2 also enhances the scalability of city-scale reconstruction through boundary-aware partition optimization, which stabilizes mesh refinement. Experiments on large-scale urban reconstruction benchmarks demonstrate that City-Mesh3R-v2 achieves higher geometric accuracy and superior mesh quality than the compared state-of-the-art city-scale reconstruction methods. The proposed frozen boundary support band enables up to a 3$\times$ speedup in mesh refinement relative to City-Mesh3R, while the complete pipeline achieves the lowest average end-to-end runtime to reconstruct the mesh surface from the sparse SFM, among the evaluated methods. These results highlight the effectiveness of City-Mesh3R-v2 for scalable, efficient, and simulation-ready city-scale 3D mesh reconstruction.