Title: Enhancing 3D Gaussian Splatting Through Point Cloud Upsampling Techniques

arXiv:2606.00450v1 Announce Type: new

Abstract:

The efficacy of 3D Gaussian Splatting (3DGS), a method for generating and rendering three-dimensional environments, is significantly contingent upon the fidelity of its initial seed points. To address this dependency, this paper investigates and assesses multiple point cloud upsampling strategies, including linear and triangular interpolation, spline-based surface reconstruction, moving least squares surface fitting, and Voronoi-based point generation. Furthermore, the study proposes a novel depth-guided point lifting technique designed to preserve geometric alignment with Structure-from-Motion (SfM) reconstructions by utilizing depth maps.

Comprehensive evaluations conducted on the Mip-NeRF360 and Replica datasets reveal that these proposed methods yield enhanced reconstruction quality across a variety of scene types. The analysis highlights that specific upsampling techniques are superior in distinct contexts: surface reconstruction algorithms prove more effective for complex, organic scenes with high detail, whereas basic interpolation methods are better suited for environments characterized by piecewise-smooth geometries. Meanwhile, the depth-guided strategy demonstrates significant potential for introducing geometry-aware points throughout the entire scene, particularly within areas lacking texture. These results offer initial practical guidance for choosing the most appropriate upsampling approach based on specific scene attributes and computational limits, thereby deepening the understanding of how point cloud initialization influences the overall quality of 3DGS.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC