ReScene framework reconstructs 3D indoor scenes with improved accuracy · arXiv paper

Researchers have developed ReScene, a new framework designed to construct simulation-ready 3D indoor scenes from multi-view captures. This method addresses limitations in existing approaches by focusing on cross-view relation fusion and physically plausible scene assembly, rather than just single-object reconstruction. ReScene utilizes a HierView component for prioritizing reconstruction views and Relation-Aware Assembly to integrate multi-frame predictions with geometric priors, resulting in a confidence-weighted scene graph. The framework achieves state-of-the-art performance on ScanNet scenes, significantly reducing Chamfer Distance and LPIPS metrics while operating faster than previous multi-view techniques. Additionally, ReScene enables the creation of a new embodied visual question answering dataset, where a fine-tuned Qwen-VL model demonstrates strong spatial reasoning capabilities. AI

IMPACT Enhances the creation of realistic 3D environments for embodied AI research and applications.

RANK_REASON The cluster contains an arXiv paper detailing a new framework for 3D scene reconstruction.

Read on arXiv cs.CV →

• arXiv
• Chamfer distance
• embodied artificial intelligence
• HierView
• lpips
• Qwen VL
• Relation-Aware Assembly
• ReScene
• SCANNET

• paper
• product

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →