Semantic-guided 3D building reconstruction from triangle meshes
Senyuan Wang, Xinyi Liu, Yongjun Zhang, Jonathan Li, Siyuan Zou, Jipeng Wu, Chuang Tao, Quan Liu, Guorong Cai
Abstract
Planar primitives tend to be incorrectly detected or incomplete in complex scenes where adhesions exist between different objects, resulting in topology errors in the reconstructed models. We propose a semantic-guided building reconstruction method known as semantic-guided reconstruction (SGR), which is capable of achieving the independence and integrity of building models in two key stages. In the first stage, the space partition is represented by a 2.5D convex cell complex and is capable of restoring planar primitives that are easily lost and can further infer the potential structural adaptivity. The second stage incorporates semantic information into a graph-cut formulation that can assist in the independent reconstruction of buildings while eliminating interference from the surrounding environment. Our experimental results confirmed that the SGR method can authentically reconstruct weakly observed surfaces. Furthermore, qualitative and quantitative evaluations show that SGR is suitable for reconstructing surfaces from insufficient data with semantic and geometric ambiguity or semantic errors and can obtain watertight models considering fidelity, integrity and time complexity.