Dual-Mode Roof Segmentation and Structured Fitting Method for Building Point Clouds

Structured roof reconstruction from point clouds is essential for 3D city modeling, rooftop photovoltaic (PV) deployment, and urban planning. However, existing model-driven, datadriven, and deep learning approaches either lack flexibility, are sensitive to noise, or rely heavily on large annotated datasets. We propose a geometry-first dual-mode framework that integrates foundation-model segmentation with classical geometric fitting. Our pipeline first isolates single-building point clouds using SAM2-guided segmentation and clustering, then applies a two-stage plane detection strategy combining robust global fitting with normal-based clustering. Roof boundaries are polygonized through angular analysis, followed by global corner consolidation and KD-Tree deduplication to ensure geometric accuracy and structural consistency. Without the need for large-scale training data, our method directly produces GIS/CAD-ready roof models. Experiments on the Building3D benchmark and a complex villa dataset demonstrate consistent improvements in plane recall, corner accuracy, and structural consistency over state-of-the-art baselines.