Para-Roof: High-quality parametric roof reconstruction via primitive proposal extraction from point cloud

3D parametric roof modeling remains a significant challenge in building reconstruction due to the geometric complexity of roofs and the noise inherent in real-world scan data. We present Para-Roof, a novel framework for high-quality parametric roof reconstruction from point clouds. Our key innovation lies in formulating the roof as a composition of geometric primitive proposals, where detected planes serve as foundational elements. To address noise-corrupted planes—a major bottleneck in prior work—we introduce a statistically driven refinement technique that optimizes plane normals and enforces geometric consistency constraints. The refined planes are then clustered into primitive proposals representing roof components. By pruning planes that disrupt proposal formation and systematically integrating viable candidates, our method achieves precise parametric reconstruction while maintaining robustness to noise. Extensive experiments demonstrate that Para-Roof outperforms state-of-the-art methods in both accuracy and computational efficiency, particularly in handling complex roof structures from real-world scans. The resulting models are CAD-ready, advancing the practicality of automated roof reconstruction for urban modeling applications, and it can also be re-edited. The source code of Pararoof is publicly available athttps://github.com/YuncongLiu-666/Pararoof.