pengsongyou/openscene

Combines vision-language models (OpenSeg/LSeg) with 3D scene geometry through multi-view feature fusion and distillation into lightweight 3D neural networks. Supports arbitrary open-vocabulary queries on indoor and outdoor scenes (ScanNet, Matterport3D, nuScenes, Replica) without requiring retraining for new concepts. Provides both a real-time interactive demo and pre-trained models for semantic segmentation, instance segmentation, and property prediction tasks.

800 stars. No commits in the last 6 months.