Robust Instance Segmentation via 3D Gaussian Denoising Constraints.

3D Gaussian Splatting (3DGS) has recently gained significant attention as an efficient representation for 3D scene modeling and photo-realistic rendering. However, achieving robust instance-level segmentation within this representation remains challenging due to inter-instance interference, noisy features, and limited semantic discriminability. These issues often lead to background artifacts, blurred boundaries, and incomplete surfaces when extracting individual instances. To address these challenges, we propose DenoGS, a global-local instance segmentation framework that explicitly models each object by integrating global semantic context with local geometric details, thereby enhancing segmentation performance. To further improve boundary accuracy and robustness, we introduce a denoising constraint that mitigates the effects of spatial noise and ambiguity. In addition, we propose a lightweight referring mechanism that enables accurate identification of target objects using a single natural language expression, supporting language-guided instance selection. Experimental results on multiple 3D scene benchmarks demonstrate that our method consistently outperforms existing approaches in both instance segmentation accuracy and referring consistency, validating the effectiveness and practicality of the proposed denoising-regularized modeling strategy and lightweight reasoning mechanism.