3D Gaussian Splatting (3DGS) has emerged as a leading method for novel-view synthesis, yet its performance is often limited in textureless regions. This limitation stems from insufficient geometric constraints, which can lead to incorrectly positioned or scaled Gaussians, manifesting as noticeable floating or blurry artifacts in rendered images. To address this issue, we propose the Geometric-Photometric Fusion Network (GP-FusionNet), a dual-stream architecture designed for robust depth refinement. The network fuses two complementary sources of information: strong geometric priors provided by ground-truth depth from our custom dataset captured with a binocular depth camera, and photometric consistency cues derived from multi-view images for fine-grained detail. By combining these inputs, GP-FusionNet generates accurate and dense depth maps. These maps provide a high-quality geometric initialization for 3D Gaussian Splatting, thereby enhancing reconstruction fidelity. We conduct extensive qualitative and quantitative evaluations on our custom dataset and selected textureless scenes from public benchmarks. Experiments show that, compared to baselines relying solely on SfM point clouds, our method significantly reduces rendering artifacts and reconstructs low-texture surfaces with higher fidelity.