A Method for 3D Building Individualization Integrating SAMPolyBuild and Multiple Spatial-Geometric Features.

Area of Science:

• Geoinformatics
• Computer Vision
• 3D Modeling

Background:

• Accurate 3D building models are crucial for urban planning and management.
• Existing methods for 3D building individualization are often manual and inefficient, or rely on deep learning with extensive pre-training and sensitivity to spatial structures.

Purpose of the Study:

• To propose a novel, efficient, and accurate method for 3D building individualization.
• To overcome the limitations of manual separation and deep learning approaches in 3D building model establishment.

Main Methods:

• Integration of SAMPolyBuild (Segment Anything Model for Polygons) with multiple spatial-geometric features for 3D building individualization.
• Utilizing SAMPolyBuild's zero-shot learning for coarse building extraction, followed by refinement using spatial-geometric features.
• Introduction of Jensen-Shannon Divergence and Earth Mover's Distance as statistical parameters for enhanced building identification.

Main Results:

• The proposed method effectively extracts individual building models from urban oblique photogrammetric 3D models.
• Achieved an F1-score of approximately 0.83 for buildings with typical spatial structures on the Semantic Urban Meshes (SUM) dataset.
• Demonstrated superior performance compared to existing methods in terms of accuracy and efficiency.

Conclusions:

• The novel method offers a feasible and effective solution for 3D building individualization.
• The integration of SAMPolyBuild and spatial-geometric features, including novel statistical parameters, significantly improves the accuracy of 3D building extraction.
• This approach provides a promising advancement for the automated generation of detailed 3D urban models.