VisGS： Physical-Visibility Aware Sparse-View Synthesis Using Gaussian Splatting

In the field of three-dimensional （3D） reconstruction， novel view synthesis plays a fundamental role that aims to generate images of a scene from unseen viewpoints given a set of input images. Recently， 3D Gaussian Splatting （3DGS） has achieved remarkable success in this task， exhibiting superior performance in different 3D applications. However， under sparse-view settings， 3DGS faces significant challenges， as its performance deteriorates due to overfitting during model training， thereby limiting its ability to generalize to unseen viewpoints. Prior studies suggest that this overfitting arises from a mismatch between model complexity and the available volume of input data. Existing approaches address this issue by reducing model complexity through random pruning， while others incorporate external depth priors as supervisory signals. However， these methods often lack principled guidance during the pruning process， leading to high variability and inconsistent improvements in synthesis performance. To overcome this limitation， we propose VisGS， a sparse-view novel view synthesis framework built upon Gaussian splatting that explicitly integrates physical visibility constraints. Specifically， we develop a physical-visibility-based pruning module that overcomes the instability of existing pruning strategies. This module deterministically evaluates whether each Gaussian primitive should be retained based on its physical visibility within the scene， thereby enabling precise and efficient elimination of redundant or erroneous primitives. Moreover， we propose a visibility-aware perturbation mechanism that injects structured perturbations into regions of high uncertainty， encouraging the model to better capture ambiguous structures and enhancing robustness under complex visibility conditions. Experimental results on benchmark datasets， including LLFF and DTU， demonstrated that， under few-shot settings， the proposed method improves the accuracy and stability of pruned regions， mitigates the uncertainty inherent in random pruning， and substantially improves the quality of novel view synthesis.