Urban Scene LOD Vectorized Modeling From Photogrammetry Meshes.

Area of Science:

• Photogrammetry and Computer Vision
• 3D Reconstruction
• Geospatial Modeling

Background:

• Urban scene modeling is complex due to scale and structural intricacy.
• Existing methods struggle with large-scale, detailed urban environments.
• Need for efficient and robust 3D urban modeling techniques.

Purpose of the Study:

• To present an efficient, multistep framework for large-scale urban scene modeling from aerial images.
• To generate compact polygon models with semantic information at various Levels of Detail (LODs).
• To develop a robust method that combines 2D and 3D information for vectorized modeling.

Main Methods:

• A segment-based modeling approach involving scene segmentation, roof contour extraction, and building modeling.
• Utilizing deep neural networks combined with 3D mesh geometric constraints for scene segmentation.
• Transforming 3D mesh data into a 2D image grid for efficient processing and modeling.

Main Results:

• The framework successfully produces compact polygon models with semantic information at different LODs.
• The 2D image grid representation leads to lower computational complexity and higher robustness.
• The method demonstrates efficiency and robustness in modeling both single buildings and large-scale urban scenes.

Conclusions:

• The proposed framework offers an efficient and robust solution for vectorized urban scene modeling.
• Combining 2D photometric and 3D geometric information is key to the algorithm's success.
• The method is flexible and does not require global prior assumptions, making it adaptable to diverse urban scenes.