ShareGS: Hole completion with sparse inputs based on reusing selected scene information

Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) are critical in numerous fields due to their high realism and efficiency. However, they face challenges in sparse input environments, where obstructions or incomplete data can cause artifacts and scene holes. Unlike prior methods primarily focused on eliminating artifacts, our research addresses the more complex issue caused by limited perspectives coverage to enhance scene completeness. We propose a method combines Gaussian homogenization and scene patches reuse to efficiently fill gaps in scenes. Firstly, a feature and scale-guided Gaussian homogenization method is proposed to increase the number of Gaussians in holes. Secondly, we utilize feature-consistent hole region selection and projection transformation strategy supervise holes completion. Finally, by leveraging unsupervised depth-color constraints, we further enhance the continuity and integrity of the scene. Our method is extensively validated on diverse datasets, demonstrating that it not only maintains scene integrity but also significantly improves novel view synthesis under constrained conditions, providing a robust solution to the complexities introduced by sparse input. • A Gaussian splatting method for hole regions completion with sparse input. • A feature and scale-guided Gaussian homogenization scheme increases Gaussian density. • Feature-consistent hole selection and projection transformation schemes guide hole completion. • A depth-color consistency regularization improves the completeness of the scene.