GeoWM model directly forecasts 3D scene geometry, reducing inference time

Researchers have introduced GeoWM, a novel world model designed for robotics and autonomous driving that directly forecasts future 3D scene geometry without relying on recursive rollouts. This approach leverages a geometry foundation model to convert RGB frames into a geometric history, which then guides a flow-matching transformer to predict geometry at specified future horizons. GeoWM also incorporates a camera-motion predictor to estimate future viewpoints, enhancing its geometric prior for forecasting. Experiments across diverse datasets demonstrate GeoWM's superior performance in predicting depth, camera pose, and 3D scene geometry, while also offering significantly reduced inference times for longer prediction horizons. AI

IMPACT This model could improve the efficiency and accuracy of 3D scene understanding for autonomous systems.

RANK_REASON Academic paper detailing a new model and its performance. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• camera-motion predictor
• flow-matching transformer
• geometry foundation model
• GeoWM
• Hugging Face
• RGB color model

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →