Radiance field rendering for virtual scene reconstruction of the built environment

Purpose—Site visualisation is of substantial importance within the field of architecture as a thorough understanding and assessment of the site informs initial design decisions. Current approaches to scanning incorporate a combined approach merging LiDAR scans with photogrammetry scans. Recent developments in computer graphics have led to radiance field rendering techniques, allowing for visually accurate virtual scene reconstruction with smaller image datasets which can be captured using any imaging device.

Design/methodology/approach—This paper examines the efficacy of emergent techniques, Neural Radiance Fields (NeRF) and 3D Gaussian splatting, by comparing results against common standard photogrammetry, specifically analysing various capabilities of NeRF and 3D Gaussian Splatting through a qualitative comparative analysis of four measures: material accuracy, interpolation competence, distance rendering and reconstruction capability.

Findings—Overall, the results indicate capabilities for emergent virtual scene reconstruction techniques to provide greater contextual awareness and accuracy for architectural/urban site analysis with easily obtained datasets, images captured from a smartphone.

Research limitations—Although the results gained have been sufficient in demonstrating capabilities of radiance field rendering techniques, hardware limitations have restricted the output render quality as GPU’s with more VRAM would allow for longer training times on NeRF, resulting in higher fidelity scene captures.