Evaluating geometric fidelity in 3D reconstruction : NeRF and Gaussian splatting
Loading...
Date
2026-09-25
Authors
Σολομωνίδη, Ελευθερία
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The digitization of cultural heritage artifacts requires reconstruction methods capable of preserving not only the overall shape of an object but also fine, localized surface details such as engravings and reliefs, which often carry much of the artifact’s significance. This thesis evaluates the ability of three reconstruction methods, Nerfacto and Instant-NGP, based on Neural Radiance Fields, and Splatfacto, based on 3D Gaussian Splatting, to preserve such fine-scale geometric detail. All three methods were implemented and trained within the Nerfstudio framework, on a datasetintroduced by the SHREC 2026 challenge, consisting of synthetic objects inspired by cultural heritage reliefs, each captured from 90 viewpoints along a helical camera trajectory.Inspection of depth renderings revealed a recurring reconstruction error common to all three methods, consisting of missing or deformed geometry that consistently aligned with saturated, overexposed regions in the input images. Comparative evaluation on the most challenging object, showed that Instant-NGP produced the leastcomplete reconstructions, with large holes through the surface, Nerfacto performed noticeably better with only minor deformations, and Splatfacto achieved the most complete and convincing results, consistent with its higher photometric quality (PSNR) across most tested objects. It was also found that a photometrically strongreconstruction does not necessarily imply high geometric accuracy, underlining the need for geometry-based evaluation.To address the identified artifact, a preprocessing technique based on the LaMa inpainting model was proposed, using a custom saturation mask generation algorithm that identifies overexposed regions on the object surface while excluding the background. Applied to Nerfacto reconstructions of an example object, thispreprocessing reduced the symmetric Chamfer distance to the ground-truth geometry and the Hausdorff distance, confirming a measurable improvement in geometric accuracy without requiring any modification to the underlying reconstruction algorithm. Code is available at https://github.com/eleftheriaa/3d-reconstruction-of-cultural-heritage
Description
Keywords
3D reconstruction, Computer vision, 3D Gaussian splatting, Nerfstudio