GeoState separates scene geometry from camera motion in 3D world models
Original title: GeoState: Learning View-Decoupled Geometry States for 3D World Modeling
Predicted states can be decoded under predicted or specified cameras without per-frame scale alignment.
GeoState is a view-decoupled geometry representation for predictive 3D world modeling. A scene tokenizer aggregates multi-view geometry-foundation-model features into a scene state that a geometry decoder queries with camera rays. An autoregressive world model predicts future scene states while camera motion is modeled separately. The authors report gains on Waymo and zero-shot generalization to KITTI and nuScenes.
- View-dependent geometry features previously mixed scene dynamics with camera variation.
- A single context-derived scale is retained throughout the rollout.
- The method supports geometrically consistent cross-view querying.
- A project page is published for the work.
From the original article
1 min read at huggingface.co
Recent 3D world models increasingly use features from geometry foundation models (GFMs) as predictive world states, benefiting from their rich geometric priors. However, the view dependence of these features entangles scene dynamics with camera-induced variation, making the resulting state evolution harder to predict.
By Hongchang Liu, Jian Liu, Yujuan Ding · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.