Neural Radiance Fields for the Digitization of Ethnographic Collections: A Comparative Analysis of State-of-the-Art and Established Methods for the 3D Documentation of Cultural Heritage Objects

Abstract

Conventional 3D documentation utilizing Structure-from-Motion (SfM) techniques has been established as an accurate and comprehensible method of cultural heritage digitization. The recent development of Neural Radiance Fields (NeRFs) is set to revolutionize the field of 3D modelling, and yet, the current appreciation of NeRFs in cultural heritage practice is still at its relative infancy. The aim of this thesis is to evaluate the performance of NeRF algorithms, implemented using Nerfstudio, as a potential alternative to the tenured method of SfM 3D visualization. This investigation addresses the quantitative and qualitative results of a comparative analysis of NeRFs as a representation of the state-of-the-art, utilizing SfM photogrammetry as reference. The quantitative results indicate varying degrees of deviation between comparable pointclouds and meshes, most of which can be attributed to the inherent differences in the implementation of each method. However, the accuracy of the 3D geometry generated by NeRF algorithms is, overall, similar to SfM references. Qualitatively, the fully trained NeRFs consistently underperform the SfM textured meshes in regard to surface details and visual quality. Ultimately, the current iterations of NeRF algorithms are not satisfactory alternatives to established SfM methods. NeRFs represent a specialized facet of the rapidly evolving field of Artificial Intelligence-assisted technologies. It is possible that within the coming years, advancements, and optimizations of the NeRF algorithm will see the method overtake current photogrammetric standards of accuracy and quality for cultural heritage digitization.