POMA-3D model learns self-supervised 3D scene understanding from point maps

Researchers have introduced POMA-3D, a novel self-supervised 3D representation model that utilizes point maps to encode 3D coordinates on a structured 2D grid. This approach allows for the transfer of 2D foundation model priors into 3D understanding through a view-to-scene alignment strategy. The model also incorporates a joint embedding-predictive architecture, POMA-JEPA, to ensure geometrically consistent features across different views. Experiments demonstrate POMA-3D's effectiveness as a backbone for various 3D tasks, including question answering and navigation, using only geometric input. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a new method for 3D scene understanding that leverages 2D priors, potentially improving performance on tasks like navigation and retrieval.

RANK_REASON This is a research paper introducing a new method for 3D scene understanding. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• POMA-3D
• POMA-JEPA
• ScenePoint
• Ye Mao

• paper
• other