Hybrid Gaussians
for Robust Open-Vocabulary
3D Segmentation

ABSTRACT

More reliable objects.
More precise segmentation.

Open-vocabulary 3D segmentation aims to localize objects in 3D scenes from free-form text queries, providing a flexible interface for language-guided scene understanding. Recent progress in vision-language models and neural scene representations has advanced this area by lifting 2D cues into 3D and enriching NeRF- or 3DGS-based scene representations with semantic or identity information. However, robust object-level segmentation remains challenging in real image sequences, where incomplete or noisy 2D supervision can destabilize multi-view identity assignment, while full-scene semantic learning can weaken object-level discriminability.

To address these issues, we introduce Hybrid Gaussians, a unified 3D representation that jointly models object association and language-aligned semantics, along with a Multi-View Object Association mechanism that combines Observation Fusion and Semantic Contrastive Learning. We further introduce Boundary Reconstruction Optimization to explicitly improve contour quality by refining local boundary structure. Experiments on LERF and 3D-OVS demonstrate strong quantitative and qualitative performance. In particular, our method achieves 59.1% mIoU on LERF, yielding a 13.43% relative gain over the baseline. These results show that jointly optimizing object association and boundary-aware refinement substantially improves open-vocabulary 3D segmentation.