3D representation of Architectural Heritage: a comparative analysis of NeRF, Gaussian Splatting, and SfM-MVS reconstructions using low-cost sensors

Abstract

This study evaluates the performance of three image-based 3D reconstruction methods—Structure-from-Motion with Multi-View Stereo (SfM-MVS), Neural Radiance Fields (NeRF), and Gaussian Splatting (GS)—for documenting Architectural Heritage (AH). Data acquisition was conducted using low-cost sensors: a drone to capture the stone portal of the external façade and a 360° camera to document the interior spaces. NeRF and GS significantly outperformed SfM-MVS in processing time, with NeRF excelling in reconstruction completeness. However, GS faced challenges with point number control, and NeRF reconstructions exhibited artifacts and noise, particularly on flat surfaces. Accuracy assessments, using TLS point clouds as a benchmark, revealed that SfM-MVS provided the highest geometric precision for the external façade, despite minor gaps in reconstruction. In contrast, NeRF and GS fell short of the accuracy required for precise geometric documentation, with NeRF exhibiting prominent artifacts in flat or poorly detailed regions. Interior reconstructions were further limited by the higher Ground Sampling Distance (GSD) caused by the technical constraints of the 360° camera and the increased capture distance for elevated areas. In conclusion, we can affirm that while NeRF and GS demonstrate strong potential for visualization due to their rendering quality and efficiency, SfM-MVS remains the most reliable method for achieving accurate geometric documentation of AH.