StruGS: Structurally consistent 3D Gaussian Splatting with targeted optimization strategies

Abstract

This paper proposes StruGS, a structural consistency-oriented optimization framework for 3D Gaussian Splatting, aiming to address the insufficient structural consistency observed in existing 3DGS methods during the representation process. This method introduces a collaborative structural optimization strategy from both the view and spatial dimensions. First, the structure-aware multi-view guidance strategy aggregates gradient signals from multiple views during training and utilizes a set of learnable structure-aware mapping parameters to guide the model to more effectively focus on structurally salient regions, thereby comprehensively enhancing the consistency of three-dimensional structural representation. Second, the structural gradient optimization balancing strategy dynamically adjusts gradients based on the depth information of each Gaussian point, ensuring a more balanced gradient optimization process across spatial regions, improving structural stability, and effectively mitigating the emergence of floater artifacts. These two strategies collaborate from the dimensions of multi-view structural guidance and spatial structural optimization balancing, enhancing structural consistency in modeling. Experimental results demonstrate that StruGS significantly improves consistency and stability in geometric structure representation and achieves high-quality novel view synthesis across multiple public datasets.