Comparison of 3D Reconstruction between Neural Radiance Fields and Structure-from-Motion-Based Photogrammetry from 360° Videos

Imagery is a standard modality of visual data capture on construction sites for documenting construction progress. Assimilating the data from multiple disjointed 2D images into a single 3D format enhances visualization and scene understanding and increases the data usability for tasks like quantity estimation and progress tracking. Two popular methods for 3D reconstruction are structure-from-motion (SfM)-based photogrammetry and neural radiance fields (NeRF), a neural network-based technique in computer vision. In this paper, we compare the spatial geometric accuracy of 3D reconstruction from 360° videos of construction sites using the SfM library called Colmap and NeRF. Our experiments show that 3D reconstruction from conventional photogrammetry is sharper than NeRF and more accurate in capturing object details and boundaries.