Sparse Input Novel View Synthesis using Gaussian Splatting for Power Distribution Network Scene

Novel view synthesis (NVS) aims to generate realistic images of a scene from viewpoints that were not part of the original observations. 3D Gaussian Splatting (3DGS) can achieve real-time rendering of unbounded 3D scenes, but this capability relies on dense training views to accurately reconstruct 3D geometric structures. When applied to power distribution network scenes with sparse views and missing perspectives, the method produces noticeable artifacts. To address this issue, we incorporate initial point cloud optimization, depth priors and a Gaussians Pruning module to handle the task of new view synthesis for sparse image inputs in power distribution network scenarios. The experimental results on sparse input power distribution network scene datasets demonstrate that the proposed approach outperforms existing methods in both visual fidelity and geometric consistency.