SynGS: Semantic-Geometric Synergistic 3D Gaussian Splatting for Scene Representation

Abstract

Accurate object-level 3D scene representation is the foundation of embodied intelligence and immersive multimedia interactions. However, existing semantic 3D Gaussian Splatting methods typically treat semantics as passive attributes of geometry. Such weakly coupled methods are highly prone to semantic aliasing at the boundaries. To address this, we propose SynGS, a novel semantic 3DGS framework based on semantic-geometric synergistic optimization. This method leverages semantic information to actively constrain the generation, evolution, and correction of Gaussian primitives. Specifically, SynGS designs a semantic-aware adaptive proliferation strategy that guides Gaussian primitives toward a reasonable spatial distribution throughout the scene through joint macro-micro scheduling. During the evolution of Gaussian primitives, semantic modulation and second-order anchor enhancement actively reshape the distribution of their geometric features. For out-of-bounds primitives, the proposed semantic correction mechanism performs pruning and reinitialization. Experiments demonstrate that SynGS achieves state-of-the-art performance on both open-vocabulary and panoptic segmentation tasks. Our method effectively eliminates boundary bleeding issues and presents sharp physical object boundaries while maintaining high-fidelity rendering.