RegionGraph: Region-Aware Graph-Based Building Reconstruction from Satellite Imagery

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Structural reconstruction is vital for scene understanding but challenged by complex real-world topologies.
• Accurate inference of spatial relationships and object layouts in scenes remains a significant computer vision problem.

Purpose of the Study:

• To propose RegionGraph, a novel method for structural reconstruction of buildings from satellite imagery.
• To address the challenges posed by complex scene structures in computer vision tasks.

Main Methods:

• RegionGraph employs a layout region graph construction and graph contraction approach.
• It introduces ConPNet (primitive estimation network) for detecting and estimating structural primitives.
• Combines structural extraction and rendering synthesis, optimizing graph structures via node-merging graph contraction.

Main Results:

• RegionGraph achieved a 4% improvement in average F1 scores across three primitive types.
• Demonstrated enhanced regional completeness and structural coherency in reconstructed building structures.
• Successfully reconstructed building structures from satellite images with improved accuracy.

Conclusions:

• RegionGraph offers an effective solution for structural reconstruction from satellite images.
• The proposed method advances the field of computer vision for scene understanding.
• The graph-based approach provides a robust framework for complex structural analysis.