SuperGSeg:基于结构高斯的开词汇3D分割
计算机视觉与模式识别
2026-01-21 v3
摘要
3D 高斯溶解最近因其高效训练和实时渲染而受到关注。虽然其标准表示主要为视图合成设计,但最近的工作将其扩展到具有语言特征的场景理解。然而,存储用于语义信息的每个高斯的额外高维特征是内存密集型的,这限制了其对复杂场景的分割和解释能力。为此,我们介绍了SuperGSeg,这是一种通过解耦分割和语言场蒸馏来促进具有连续性、上下文感知的分层场景表示的新方法。SuperGSeg首先利用可获取的2D掩码,通过神经3D高斯学习几何、实例和分层分割特征。随后,这些特征被用于创建稀疏的\acrlong{superg}集合。\acrlong{superg} facilitate the lifting and distillation of 2D language features into 3D space. They enable hierarchical scene understanding with high-dimensional language feature rendering at moderate GPU memory costs. Extensive experiments demonstrate that SuperGSeg achieves remarkable performance on both open-vocabulary object selection and semantic segmentation tasks.
引用
@article{arxiv.2412.10231,
title = {SuperGSeg: Open-Vocabulary 3D Segmentation with Structured Super-Gaussians},
author = {Siyun Liang and Sen Wang and Kunyi Li and Michael Niemeyer and Stefano Gasperini and Hendrik P. A. Lensch and Nassir Navab and Federico Tombari},
journal= {arXiv preprint arXiv:2412.10231},
year = {2026}
}
备注
13 pages, 8 figures. Project page: supergseg.github.io