PD-GS: Prior-based Decoupled Optimization for Unbounded Scene Gaussian Splatting

Abstract

Prized for its high-quality real-time rendering, 3D Gaussian Splatting is a key novel view synthesis technology for fields such as virtual reality and autonomous driving, where it is widely used for outdoor scene reconstruction. However, extensive unbounded and texture-scarce regions in outdoor scenes, such as the sky, pose challenges for 3DGS modeling due to the lack of initial point clouds, severely limiting rendering quality. To address this issue, this paper proposes a novel optimization approach for 3DGS in unbounded scenes. We first reveal the strong coupling effects between the sky and foreground during optimization and subsequently design a prior-based decoupled optimization strategy. This method is primarily implemented through two key components: prior-based skybox initialization (PBSI) and sphere-aligned skybox densification (SASD). PBSI employs semantic cues to effectively eliminate the coupling effects. SASD introduces a spherical alignment constraint to enable high-fidelity optimization of the sky regions. Experimental results demonstrate that the proposed method can effectively eliminate the aforementioned coupling effects.