A Family of Approaches for Full 3D Reconstruction of Objects with Complex Surface Reflectance
A Family of Approaches for Full 3D Reconstruction of Objects with Complex Surface Reflectance
复制标题
DOI：
10.1007/s11263-023-01795-w
复制
发表时间：
2023-06
影响因子：
19.5
通讯作者：
Gianmarco Addari;Jean-Yves Guillemaut
中科院分区：
文献类型：
--
作者：
Gianmarco Addari;Jean-Yves Guillemaut
一般场景的3D重建仍然是一个开放的挑战，目前的技术通常依赖于对场景表面反射率的假设，这限制了可以建模的对象的范围。亥姆霍兹立体视觉提供了一个吸引人的框架，使建模过程与表面反射率无关。然而，以前的配方几乎完全局限于2.5D建模。为了解决这一差距，本文介绍了一系列利用亥姆霍兹互易来产生具有任意未知反射率的物体的完整3D模型的重建方法。这包括一种基于融合（正交或透视）视图相关重建的方法，一种优化体素网格内表面位置的体积方法，以及一种优化给定网格拓扑的顶点位置的基于网格的公式。在合成数据集和真实数据集上对贡献的方法进行了评估，包括与本文一起公开发布的新型全3D数据集，并与广泛的竞争方法进行了实验比较。结果表明，不同的方法的好处和他们的能力，以实现高质量的全三维重建的复杂对象。
3D reconstruction of general scenes remains an open challenge with current techniques often reliant on assumptions on the scene’s surface reflectance, which restrict the range of objects that can be modelled. Helmholtz Stereopsis offers an appealing framework to make the modelling process agnostic to surface reflectance. However, previous formulations have been almost exclusively limited to 2.5D modelling. To address this gap, this paper introduces a family of reconstruction approaches that exploit Helmholtz reciprocity to produce complete 3D models of objects with arbitrary unknown reflectance. This includes an approach based on the fusion of (orthographic or perspective) view-dependent reconstructions, a volumetric approach optimising surface location within a voxel grid, and a mesh-based formulation optimising vertices positions of a given mesh topology. The contributed approaches are evaluated on synthetic and real datasets, including novel full 3D datasets publicly released with this paper, with experimental comparison against a wide range of competing methods. Results demonstrate the benefits of the different approaches and their abilities to achieve high quality full 3D reconstructions of complex objects.