Abstract
3D Gaussian Splatting has emerged as an efficient photorealistic novel viewsynthesis method. However, its reliance on sparse Structure-from-Motion (SfM)point clouds often limits scene reconstruction quality. To address thelimitation, this paper proposes a novel 3D reconstruction framework, GaussianProcesses enhanced Gaussian Splatting (GP-GS), in which a multi-output GaussianProcess model is developed to enable adaptive and uncertainty-guideddensification of sparse SfM point clouds. Specifically, we propose a dynamicsampling and filtering pipeline that adaptively expands the SfM point clouds byleveraging GP-based predictions to infer new candidate points from the input 2Dpixels and depth maps. The pipeline utilizes uncertainty estimates to guide thepruning of high-variance predictions, ensuring geometric consistency andenabling the generation of dense point clouds. These densified point cloudsprovide high-quality initial 3D Gaussians, enhancing reconstructionperformance. Extensive experiments conducted on synthetic and real-worlddatasets across various scales validate the effectiveness and practicality ofthe proposed framework.