Browse Teach Curate

Generalizable Open-Vocabulary Semantic Segmentation of 3D Gaussian Splatting Scenes

OVGaussian: a framework that predicts an open-vocabulary semantic property for each 3D Gaussian point with a 3D neural network and renders it into multi-view-consistent 2D semantic maps (Generalizable Semantic Rasterization), trained with Cross-modal Consistency Learning on the SegGaussian dataset so that open-vocabulary querying generalizes across novel 3DGS scenes rather than being confined to a single trained scene.

Rosie Text

2501.00326

OVGaussian is a generalizable open-vocabulary 3D semantic segmentation framework built on 3D Gaussian Splatting (3DGS). Prior lift-and-adapt methods project 2D vision-language features (e.g. CLIP) on…