Visual Hull Series: A Coarse‐to‐Fine Low‐Poly Meshing Method for Urban Buildings

In recent years, multi‐view stereo (MVS) methods have been gaining increasing popularity in the 3D reconstruction of large‐scale urban scenes. However, due to the very high element count, models produced by the MVS pipeline are usually difficult to be directly used in 3D applications. In this paper, we propose a novel coarse‐to‐fine algorithm to simplify high‐poly urban building models reconstructed through the MVS pipeline, while highly preserving main visual features under user‐specified rendering precision. Visual hull is a concept that was originally proposed for 3D reconstruction from images, which captures the silhouette features of an object. Employing it as an approximation operator for arbitrary geometry, we introduce the novel visual hull series (VHS) . Similar to the concept of series expansion in algebra, a ‐order VHS decomposes the input model into visual hull terms and a remainder term . If ignore the remainder term and combine all the visual hull terms, we will transform the VHS expansion to an approximating shape of the input model. However, there can be many different VHS expansions for an object. To determine the optimal expansion, our algorithm starts from the 0th‐order expansion, progressively appends new visual hull terms and simultaneously updates the remainder term. Such a process is equivalent to constructing the from coarse to fine. To ensure the visual similarity between the simplified mesh and the input model, we propose an image‐based error metric to guide this process. The experimental results indicate that the low‐poly meshes generated by our algorithm exhibit high visual fidelity to the input model, validating the effectiveness of the proposed method.