The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B2-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2021, 565–570, 2021
https://doi.org/10.5194/isprs-archives-XLIII-B2-2021-565-2021
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2021, 565–570, 2021
https://doi.org/10.5194/isprs-archives-XLIII-B2-2021-565-2021

  28 Jun 2021

28 Jun 2021

A COMPARISON BETWEEN 3D RECONSTRUCTION USING NERF NEURAL NETWORKS AND MVS ALGORITHMS ON CULTURAL HERITAGE IMAGES

F. Condorelli1, F. Rinaudo1, F. Salvadore2, and S. Tagliaventi2 F. Condorelli et al.
  • 1DAD, Department of Architecture and Design, Politecnico di Torino, Italy
  • 2CINECA – HPC Department, Rome, Italy

Keywords: NeRF, Neural Networks, 3D Reconstruction, Photogrammetry, MVS algorithms, Cultural Heritage, Metric Quality Assessment

Abstract. In this research, an innovative comparison between 3D reconstructions obtained by means of Artificial Intelligence, in particular NeRF Neural Networks, and by Structure-from-Motion (SfM) and Multi-View-Stereo (MVS) open-source algorithms is proposed. The 3D reconstruction comparison is performed on two test cases, one of cultural interest, one useful only for technical discussion. It is known that the approaches are traditionally used with different objectives and in different contexts but they can however also be used with similar purpose, i.e., 3D reconstruction. In particular, we were interested in evaluating how NeRF reconstructions are accurate from a metric point of view and how the models obtained from the application of NeRF differ from the model obtained from the classical photogrammetry. By analyzing the results in the considered test cases, we show how NeRF networks, although computationally demanding, can be an interesting alternative or complementary methodology, especially in cases where classical photogrammetric techniques do not allow satisfactory results to be achieved. It is therefore suggested to expand efforts in this direction by exploiting, for example, the numerous improvement proposals of the original NeRF network.