DHGS: Depth-Regularized Homogeneous 3D Gaussian Splatting for Real-Time Rendering

Representing 3D scenes from multi-view images remains a core challenge in computer vision and graphics, requiring balanced optimization between photorealistic rendering and geometric accuracy. Existing methods often prioritize image quality over structural fidelity, leading to suboptimal trade-offs due to limitations in geometric supervision and Cartesian coordinate constraints. We propose DHGS, a framework integrating 3D Gaussian Splatting (3DGS) with depth-supervised learning through homogeneous coordinate embedding and adaptive monocular-SfM depth fusion. By mapping Gaussian primitives into projective space, our method resolves scale ambiguity in distant-view rendering while enhancing local geometric details via confidence-aware loss weighting. Experiments demonstrate DHGS improves PSNR by 1.82 dB and reduces geometric error by 37.6 % compared to 3DGS, maintaining real-time efficiency (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$200+$</tex> FPS). These advancements enable practical applications in VR large-space exploration, where coherent distant scene details during rapid viewpoint changes mitigate motion sickness.