Abstract: [Purposes] 3D Gaussian splatting (3DGS)has emerged as a new paradigm for efficient and photorealistic 3D reconstruction.Under resource-constrained conditions,planning a limited number of supplementary viewpoints is an effective way to mitigate scene uncertainty and improve reconstruction quality.However,existing studies lack global scene guidance and often restrict uncertainty estimation to single-view observations,which limits global batch view planning and constrains its applicability in large-scale automated data acquisition and reconstruction. [Methods] To address these challenges,this paper introduces Fisher information into 3D Gaussian splatting and develops a scene-level uncertainty modeling approach that combines color and depth rendering gradients for comprehensive uncertainty estimation.A globally uncertainty-guided batch view planning strategy is then proposed to achieve efficient large-scale viewpoint optimization in a single step. [Findings] Experiments on several representative architectural scenes demonstrate that the proposed method significantly reduces data acquisition requirements while improving reconstruction completeness and efficiency,maximizing reconstruction quality with minimal resource consumption. [Conclusions] This approach enhances large-scale,high-fidelity 3D reconstruction under resource constraints and provides technical support for real-time environmental perception and digital twin applications.

Key words: 3D reconstruction, uncertainty, 3D Gaussian splatting, Fisher information, view planning