Density-Regularized Robust Loss for 3D Gaussian Splatting in High-Fidelity Heritage

3D Gaussian Splatting (3DGS) has recently emerged as a transformative technique for photorealistic rendering in digital heritage preservation, offering a promising alternative to traditional Neural Radiance Fields (NeRF) due to its real-time rendering capabilities and explicit representation. However, applying 3DGS to heritage scenarios, which often involve complex architectural geometries, intricate textures, and challenging capture conditions, reveals critical limitations. Specifically, 3DGS frequently suffers from artifacts such as needle-like Gaussian protrusions and floating ”pseudo-surfaces” in unconstrained regions, significantly compromising the geometric fidelity essential for scholarly documentation and virtual restoration. To address these challenges, we propose the Density-Regularized Robust Loss (DRRL), a novel optimization framework designed to enhance the structural integrity and visual quality of 3DGS in heritage contexts. Our approach introduces two core components: a density-based regularization term that penalizes abnormally high-density Gaussians, effectively mitigating needle-like artifacts and suppressing floating outliers, and a robust loss function that dynamically adapts to pixel-wise reconstruction errors. This robust loss mechanism prevents over-fitting to transient occlusions and noisy data, which are common in heritage site imaging, thereby preserving high-frequency textural details without introducing structural distortions. We evaluate our method on different datasets, ranging from intricate statues to architectural scenes. Experimental results demonstrate that DRRL outperforms state-of-the-art 3DGS baselines, delivering superior rendering quality and more accurate geometric reconstruction. By effectively eliminating artifacts and ensuring high-fidelity reconstruction, our method provides a reliable and efficient solution for digital heritage applications.