DP-GS: Depth-prior & Perception-guided Gaussian Splatting for Sparse-view Novel View Synthesis

Sparse-view novel view synthesis (NVS) is crucial for practical applications such as AR/VR, robotics, and largescale scene reconstruction, where capturing dense multi-view images is often impractical. 3D Gaussian Splatting (3DGS) offers explicit and efficient scene representations for real-time NVS. However, its performance significantly degrades under sparseview settings, resulting in incomplete geometry and severe texture artifacts. To overcome these limitations, we propose DP-GS (Depth-prior & Perception-guided Gaussian Splatting), a unified framework designed for sparse-view NVS. DP-GS integrates depth regularization derived from monocular depth estimation to improve geometric completeness. Moreover, we introduce a point-level dropout mechanism to suppress unreliable points, and integrate perceptual optimization guided by diffusion model to enhance texture fidelity and structural consistency. Extensive experiments on LLFF and Mip-NeRF360 datasets demonstrate that DP-GS consistently outperforms existing 3DGS methods under sparse view settings.