ETD

• 3D gaussian splatting
• 3D reconstruction
• high dinamic range imaging
• no reference metric
• quick acquisition
• user study

Recent advances in novel view synthesis have transformed 3D reconstruction, making image-based acquisition more accessible and efficient while preserving high visual quality. Among these methods, 3D Gaussian Splatting (3DGS) stands out for combining state-of-the-art rendering quality with real-time performance. However, limited attention has been devoted to complete reconstruction pipelines that preserve the efficiency goals of these methods from acquisition to evaluation.

This thesis proposes an efficient end-to-end workflow for 3DGS-based reconstruction. First, High Dynamic Range (HDR) imaging is integrated into the pipeline to better capture luminance variations and preserve details in challenging lighting conditions. A dedicated HDR dataset was acquired, the impact of tone mapping on reconstruction quality was analyzed, and a consistent multi-view tone mapping strategy was introduced to ensure coherence across viewpoints.

Second, a rapid multi-camera acquisition system based on three synchronized cameras was designed to accelerate data collection while maintaining reliable multi-view coverage. The system was validated in cultural heritage and biomedical case studies.

Finally, the thesis investigates no-reference quality assessment methods for 3DGS reconstructions. A user study involving 32 participants enabled the development of a perceptually inspired metric capable of estimating reconstruction quality without ground truth data. Overall, the proposed modular pipeline jointly optimizes acquisition, reconstruction, and evaluation for practical real-world applications.

In anni recenti i progressi della novel view synthesis hanno trasformato la ricostruzione 3D, rendendo l’acquisizione da immagini più accessibile ed efficiente mantenendo alta qualità visiva. Tra i metodi, 3D Gaussian Splatting (3DGS) si distingue per rendering avanzato e prestazioni in tempo reale. Rimane però limitata l’attenzione a pipeline complete che preservino tali obiettivi dall’acquisizione alla valutazione.

La tesi propone una pipeline end-to-end per 3DGS. Integra imaging High Dynamic Range (HDR) per catturare meglio variazioni di luminanza e dettagli in condizioni difficili. È stato acquisito un dataset HDR, analizzato l’impatto del tone mapping e introdotta una strategia multi-view coerente per garantire consistenza tra viste.

In secondo luogo è stato progettato un sistema di acquisizione multi-camera rapido con tre camere sincronizzate, per ridurre i tempi di raccolta dati mantenendo adeguata copertura multi-vista. Il sistema è stato validato in ambito beni culturali e biomedico.

Infine si studiano metodi no-reference per valutare la qualità delle ricostruzioni 3DGS. Uno studio con 32 partecipanti ha permesso di sviluppare una metrica percettiva capace di stimare la qualità senza ground truth. La pipeline proposta ottimizza acquisizione, ricostruzione e valutazione per applicazioni reali.