Browsing by Author "Pierdicca, Roberto"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Comparing MVS and Gaussian Splatting for the 3D Reconstruction of Reflective and Texture-less Cultural Heritage Artifacts
    (The Eurographics Association, 2025) Clini, Paolo; Pierdicca, Roberto; Nespeca, Romina; Angeloni, Renato; Coppetta, Laura; Campana, Stefano; Ferdani, Daniele; Graf, Holger; Guidi, Gabriele; Hegarty, Zackary; Pescarin, Sofia; Remondino, Fabio
    This study presents a comparative analysis of Multi-View Stereo (MVS) and 3D Gaussian Splatting (GS) for the three-dimensional reconstruction of cultural heritage artifacts characterized by reflective and texture-less surfaces, conditions that traditionally challenge image-based modeling techniques. Two case studies, a ceramic and a bronze head, were documented through controlled photographic acquisition and processed using both methods. Laser scanning served as a geometric benchmark for quantitative evaluation. Results demonstrate that GS provides more spatially homogeneous reconstructions and enhanced visual coherence, particularly in regions affected by specular highlights and low surface texture. Visual comparisons further highlight the potential of GS to deliver smooth, photorealistic renderings through its volumetric, view-dependent representation.
  • Thumbnail Image
    Item
    Enhancing Cultural Heritage with Generative AI: A Comparative Framework for the Evaluation of 3D Model Accuracy and Visual Fidelity
    (The Eurographics Association, 2025) Balloni, Emanuele; Paolanti, Marina; Uggeri, Jacopo; Zingaretti, Primo; Pierdicca, Roberto; Campana, Stefano; Ferdani, Daniele; Graf, Holger; Guidi, Gabriele; Hegarty, Zackary; Pescarin, Sofia; Remondino, Fabio
    The digitization of Cultural Heritage (CH) has become a vital tool for preservation and dissemination, with 3D reconstruction playing a key role in capturing intricate geometries and visual details of artifacts. While traditional methods like photogrammetry and laser scanning are effective, they often involve labor-intensive processes and struggle with complex material properties. Recent advancements in Generative AI (GenAI), particularly Large Reconstruction Models (LRMs) such as TRELLIS, offer promising alternatives for 3D generation. However, their application in CH remains underexplored. This paper introduces a novel comparative framework to evaluate the accuracy and visual fidelity of 3D GenAI models in the CH domain. Focusing on TRELLIS, the framework assesses single-view and multi-view 3D generation across five diverse CH scenes, employing both 2D (PSNR, SSIM, LPIPS) and 3D (Chamfer Distance, F-score, Accuracy) metrics. Results demonstrate superior performance for individual artifacts (e.g., Minareto, Greek Vase) compared to complex architectural scenes, with multi-view generation consistently outperforming single-view approaches. The study highlights the potential of GenAI for CH preservation while identifying challenges in large-scale reconstructions, paving the way for future hybrid methodologies and sparse-view optimizations.
  • Thumbnail Image
    Item
    Prompting Meaning: Optimizing Prompt Engineering for Architectural Point Cloud Interpretation
    (The Eurographics Association, 2025) Paolanti, Marina; Muralikrishna, Nikhil; Gorgoglione, Lucrezia; Pierdicca, Roberto; Campana, Stefano; Ferdani, Daniele; Graf, Holger; Guidi, Gabriele; Hegarty, Zackary; Pescarin, Sofia; Remondino, Fabio
    Three-dimensional point cloud visualisation is essential for preserving and analysing built heritage by providing detailed insights into architectural forms and spatial configurations. Although human perception naturally integrates visual, spatial and contextual information, AI systems have yet to match this interpretive ability, particularly about 3D point clouds. This gap in interpretation highlights the need for AI approaches that process 3D data not only geometrically but also semantically. To address this challenge, the 3D.LLM project is exploring how combining point clouds with large language models (LLMs) can improve spatial and linguistic understanding. This paper presents a prompt engineering strategy developed as part of the 3D.LLM project to improve the semantic interpretation of architectural point clouds. By linking spatial attributes to language-based reasoning, LLMs are employed to generate richer and more accurate descriptions of cultural heritage environments. Unlike conventional geometric segmentation approaches, which often fail to capture architectural nuances, this system enables a spatially aware and flexible interpretation of 3D data. To refine the AI outputs and ensure spatial precision, domain-specific benchmarks such as ArCH and Objaverse XL have been employed. Preliminary findings suggest that prompt engineering significantly improves interpretability, descriptive accuracy and contextual depth, outperforming traditional automated methods. Beyond improving accessibility to architectural heritage information, this approach encourages interdisciplinary collaboration by making complex 3D structures more accessible and useful to scholars, conservators and a wider audience.