SemanticMesh: parameterized fusion of semantic components for photogrammetric meshes

The façades of photogrammetric building mesh models frequently lack detailed semantics, such as windows, doors, and balconies. To address this, we introduce SemanticMesh, a methodology designed to enrich façades with detailed semantics by integrating discrete components seamlessly. The process begins by transforming a localized area of the building surface into a 2D planar mesh using geodesic vectors, based on the component poses. This is followed by the application of 2D constrained triangulation to map the component boundaries onto the parameterized plane, after which the mesh is reconverted into 3D space. To achieve a smooth and aesthetically pleasing integration, we employ a Laplacian mesh deformation technique along the 3D embedding boundary. Our experiments across three distinct datasets – featuring near-planar, non-planar, and noisy surfaces – demonstrate that SemanticMesh provides superior modeling outcomes compared to conventional approaches.