A monocular Gaussian splatting SLAM framework for cultural heritage digitization

Abstract High-fidelity 3D documentation of outdoor cultural heritage is essential for conservation and digital preservation, yet remains challenging under uncontrolled conditions with monocular inputs. This paper presents a monocular Gaussian Splatting SLAM with viewpoint-aware optimization framework tailored for large-scale heritage digitization. The method integrates pointmap-based pose initialization with a viewpoint-aware optimization strategy to mitigate scale drift and maintain geometric consistency under wide-baseline observations and complex illumination. Leveraging the explicit representation and efficient optimization of 3D Gaussian splatting, the framework achieves robust performance in large outdoor environments. Experiments on real-world heritage datasets and public benchmarks (Tanks and Temples) demonstrate improved trajectory accuracy and rendering quality over state-of-the-art GS-SLAM methods. Beyond performance gains, the approach supports structurally coherent and interpretable digital representations, contributing to more reliable documentation and understanding of cultural heritage.