Evaluation of 3D Reconstruction of Non-collaborative Surfaces using Neural Radiance Fields

Abstract

This paper investigates the effectiveness of Neural Radiance Fields (NeRF) in reconstructing the three-dimensional models of non-collaborative surfaces, with a specific focus on transparent and reflective materials. These surfaces often pose significant challenges to traditional photogrammetric methods due to their optical complexity. The study compares the accuracy of NeRF with that of conventional photogrammetric techniques based on Multi-View Stereo (MVS). Two objects, a plastic and an aluminium water bottle, were photographed with a smartphone and processed in both Agisoft Metashape and NefStudio to generate 3D models. The reconstructed models were evaluated against reference point clouds obtained through laser scanning. Accuracy assessment was conducted using the M3C2 analysis in CloudCompare, in which NeRF achieved mean distance errors of 0.650 mm and 0.153 mm for transparent and reflective surfaces, respectively. These values were lower than those obtained from MVS, which recorded 0.833 mm and 0.088 mm, respectively. The results indicate that while photogrammetry yields reliable outcomes for textured surfaces, NeRF demonstrates improved performance in modelling complex reflective and transparent geometries. These findings support the potential of NeRF as a practical alternative or complementary approach to traditional photogrammetry in scenarios involving challenging surface characteristics.