AI-Based 3D Modeling Strategies for Civil Infrastructure: Quantitative Assessment of NeRF and Photogrammetry

Area of Science:

• Civil Engineering
• Computer Vision
• Geomatics

Background:

• Three-dimensional (3D) modeling is crucial for infrastructure analysis, supervision, and planning.
• Image acquisition for 3D modeling can be operationally complex, increasing fieldwork time and costs.
• Evaluating reconstruction techniques under data constraints is vital for efficient infrastructure monitoring.

Purpose of the Study:

• To comparatively assess Neural Radiance Fields (NeRFs) and digital photogrammetry for 3D infrastructure modeling.
• To analyze the performance of NeRF and photogrammetry under varying data availability and image acquisition conditions.
• To develop a framework for selecting appropriate 3D modeling technologies based on engineering objectives and resource constraints.

Main Methods:

• A real-world case study of a terrace at the Pontificia Universidad Católica de Valparaíso was used.
• Both NeRF and digital photogrammetry were evaluated with progressively reduced image datasets.
• Performance metrics included visual quality, point cloud density, geometric accuracy, and processing time.

Main Results:

• NeRF provided rapid, photorealistic 3D reconstructions with fewer images.
• Digital photogrammetry demonstrated superior metric accuracy and structural completeness.
• NeRF's efficiency with reduced data supports infrastructure monitoring under variable conditions.

Conclusions:

• NeRF excels in speed and visual quality with limited data, while photogrammetry leads in geometric precision.
• An application-oriented framework and hybrid workflows can guide the selection of 3D modeling technologies.
• AI-based reconstruction methods like NeRF can enhance digital workflows for infrastructure monitoring with constrained data.