Assessment of 3D Gaussian Splatting reconstruction quality from still frames versus 360-degree video acquisitions: the Osservanza complex in Imola

This research is conducted within the framework of the “Osservanza Innovation Park” urban regeneration project, funded by the Italian National Recovery and Resilience Plan (PNRR) under the Integrated Urban Plans initiative (official website: https://www.conami.it/pnrr-conami/piani-integrati-osservanza/). The project aims to revitalize Imola’s former psychiatric hospital into a hub for education, innovation, research, and social inclusion. In particular, we present our contribution within this framework, which consists of the survey and creation of 3D Gaussian Splatting (3DGS) (Kerbl et al., 2023) of a selection of the external spaces of the Osservanza complex in Imola. As a landscape characterized by intricate vegetation, historical architectural elements, and new buildings, the area serves as an ideal testing ground for advanced digitization techniques (Chen & Wang, 2024). The primary objectives of this study are to develop and test a cost-effective survey and digitization workflow and tools, as well as to evaluate and compare two distinct field acquisition methodologies: high-resolution photography utilizing a standard mirrorless digital camera, and continuous environmental video capture utilizing a 360-degree camera (Zhang et al., 2025). The choice of using 3D Gaussian splats instead of traditional photogrammetric mesh models or 360 panoramic static images comes from the fact that 3D splats enable users to have more immersive interactive navigation on a variety of devices (Kwon & Yu, 2025), including mid-end smartphones, and provide better realism for vegetation-heavy, detailed, wide scenes and in the presence of highly specular materials. The workflow based on static photos involves the systematic, manual capture of overlapping high-resolution images. While this approach reliably yields maximum visual fidelity and improved camera tracking during the initial Structure-from-Motion (SfM) alignment phase (Arrigoni, 2025), it is notably time-consuming in the field and thus can struggle with the dynamic lighting changes inherent to outdoor environments. Conversely, the 360-degree video methodology proposes a rapid-acquisition alternative. By extracting sequential frames from a continuous spherical video recorded while traversing the site, field survey time is drastically reduced. Furthermore, the omnidirectional nature of the 360-degree lens ensures comprehensive spatial coverage, possibly reducing blind spots and capturing a holistic view of complex environmental geometries in a single pass. This comparative analysis evaluates both approaches across several critical metrics: time efficiency during field acquisition, computational demands during alignment and Gaussian Splat training, accuracy of the photogrammetric camera alignment, and the ultimate visual fidelity of the rendered scene. Preliminary results indicate that using static traditional pictures better resolves finer detail, providing overall increased perceived sharpness in the final rendered images, while the 360-degree video workflow provides a more efficient pipeline for large-scale landscapes, improving consistency in areas characterized by dynamic light changes. Ultimately, this study defines a cost-effective workflow for landscape heritage, supporting the democratization of digital surveying tools and empowering institutions to create accessible, high-fidelity virtual restitution of reality. The postprocessing of the pictures for the creation of the Gaussian splats was performed entirely in free or open source software packages (e.g., Reality Scan and Licht Feld). Furthermore, a novel open desktop application for 360-degree frame elaboration was developed through Vibe-Coding.