PACG: Prior-Guided Sparse Surface Reconstruction With Complementary Densification-Pruning Based on Gaussian Splatting

Gaussian Splatting has emerged as a new trend in surface reconstruction methods due to its camera-level novel view synthesis and accurate reconstruction. However, existing approaches mainly rely on dense views, often suffering from overfitting of Gaussian primitives that generate catastrophic floaters in sparse-view scenes, significantly degrading reconstruction accuracy. In this work, we propose PACG, a novel sparse-view surface reconstruction framework that integrates prior models with complementary densification-pruning. Specifically, PACG leverages the robust priors of Diffusion model and Feedforward model to initialize reconstruction. A coarse-to-fine geometry-aware loss ensures optimization stability, while gradient-uncollision densification and contribution-based pruning are employed during optimization to suppress floaters. PACG demonstrates superior performance over existing methods, achieving SOTA results on widely used DTU, Replica and BlendedMVS datasets.