<p>Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) achieve compelling results under dense-view supervision, yet their performance degrades significantly in sparse-view settings due to insufficient cross-view constraints and unreliable geometric initialization. Although depth- or epipolar-guided methods partially mitigate this ambiguity, appearance-based correspondence remains fragile in textureless or repetitive regions. Furthermore, static initialization often leads to persistent geometric holes during optimization. To address these challenges, we propose SARG-GS, a geometry-driven 3DGS framework tailored for sparse-view scenarios, comprising a Semantic Augmented Epipolar Fusion (SAEF) module and a Residual Guided Reprojection Compensation (RRC) module. Specifically, SAEF integrates high-level semantic features into an epipolar-constrained matching formulation, employing sub-pixel Soft-Argmax refinement to generate dense, geometrically consistent initialization points. RRC then utilizes residual-guided 3D reprojection to dynamically inject and adjust Gaussian primitives, facilitating continuous geometric correction throughout training . Extensive experiments on the LLFF, Blender, DTU, and Mip-NeRF360 datasets demonstrate that SARG-GS achieves superior structural completeness and rendering fidelity with as few as three input views, significantly reducing floating artifacts and structural holes compared to prior sparse-view baselines.</p>