GeoState: Learning View-Decoupled Geometry States for 3D World Modeling
GeoState decouples scene geometry from camera motion for predictive 3D world modeling.
GeoState proposes a view-decoupled geometry representation for predictive 3D world modeling, separating scene dynamics from camera-induced variation. A scene tokenizer aggregates multi-view geometry-foundation features, a decoder queries them with camera rays, and an autoregressive model predicts future states while camera motion is modeled separately. The method retains one context-derived scale without per-frame alignment. Experiments report better results than prior 3D world models on Waymo, consistent cross-view querying, and zero-shot generalization to KITTI and nuScenes.
- GeoState separates scene-state evolution from camera-dependent observations.
- A tokenizer aggregates multi-view geometry features into one scene state.
- Future states remain decodable under predicted or specified cameras.
- It outperforms prior models on Waymo and generalizes to KITTI and nuScenes.
From the original article
1 min read at arxiv.orgRecent 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 · arxiv.org
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.