Based on depth-aware and scale-adaptive 3D Gaussian splatting for sparse-view reconstruction

3D Gaussian Splatting (3DGS) has outperformed Neural Radiance Field methods in novel view synthesis by offering lower computational costs and high-quality real-time rendering. However, 3DGS heavily depends on continuous viewpoints, and its synthesis quality significantly degrades under sparse input conditions. To address this limitation, this study proposes a depth-guided geometric optimization framework to enhance novel view synthesis in few-shot scenarios. Initially, monocular depth estimation is used to generate an initial depth map, which is subsequently refined and aligned through nonlinear depth transformation and geometric priors. Concurrently, the scale of 2D Gaussian projections is dynamically adjusted, enabling the Gaussian distributions to progressively conform to the geometric structure implied by the depth prior. Finally, local consistency regularization is applied in regions with missing textures or geometric inconsistencies. Experimental results demonstrate that the proposed method effectively enhances novel view synthesis quality under limited input conditions.