Efficient 3D reconstruction of heritage monuments using symmetry-aware part-based modeling

Abstract

3D reconstruction of heritage monuments is crucial for preservation and analysis, yet traditional methods are resource-intensive. This study introduces a cost-effective solution tailored for low- and middle-income countries, leveraging internet images contributed by visitors. Focusing on Indo-Islamic monuments, which exhibit four-way symmetry and modular architectural parts, our approach utilizes a simplified orthographic variant of the voxel carving algorithm. This method significantly reduces computational requirements and allows satisfactory reconstruction from a single fronto-parallel view, which can be refined with additional views. Our final 3D models demonstrate high accuracy, completeness, regularity, and smoothness, respectively, achieving a Chamfer-L2 distance of 0.002 (0.001), a voxel IoU of 0.799 (0.836), a nearest neighbor coefficient of variation of 0.418 (0.369), and a mean curvature of 0.001 (0.001), based on single view (two views), when evaluated against a reference CAD model of the Taj Mahal that serves as ground truth for all competing methods. The source code for this work is publicly available at https://github.com/BarnitaSharma/Part-based-3D-Reconstruction and is permanently archived on Zenodo (DOI: https://doi.org/10.5281/zenodo.18006077).