3D modeling of cultural spaces using Artificial Intelligence: from rapid and low resolution video capture to high performance reconstruction

This research explores the use of low-resolution monocular and 360° video for 3D reconstruction in the context of cultural heritage documentation. The proposed methodology relies on GLOMAP, an efficient global Structure-from-Motion pipeline, to generate sparse point clouds from consumer-grade video data. To enhance the visual quality and address limitations due to low resolution and incomplete depth information, the workflow integrates Artificial Intelligence techniques, specifically Neural Radiance Fields and 3D Gaussian Splatting. These methods improve texture continuity, reduce visual artifacts, and provide photorealistic rendering capabilities, even in challenging conditions such as the data examined. The pipeline was tested on real-world case studies in South Tyrol, including architectural documentation, rapid surveying of temporary museum exhibits, and immersive environment creation for educational videogames. The results demonstrate that while laser scanning and traditional photogrammetry remain more metrically accurate, the proposed solution offers a fast, cost-effective, and visually convincing alternative, especially suitable for dissemination and interactive applications. Challenges related to geometric precision and computational cost were identified, suggesting future research directions aimed at optimizing AI integration and depth estimation. Overall, this work highlights the potential of combining SfM and neural rendering for accessible and scalable heritage digitization.