Line-Based 3D Building Abstraction and Polygonal Surface Reconstruction From Images

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geometric Modeling

Background:

• Traditional 3D reconstruction methods struggle with textureless objects and repetitive patterns.
• Feature-point based methods fail due to lack of distinct texture or ambiguous matches.
• Multi-view stereo approaches are computationally intensive.

Purpose of the Study:

• To present a new framework for reconstructing 3D surfaces of man-made structures from multi-view images.
• To leverage line segments as a fundamental geometric primitive for improved reconstruction.
• To address challenges posed by textureless objects and limited computational resources.

Main Methods:

• A multi-resolution line segment detector extracts 2D line segments from images.
• An improved Line3D++ algorithm constructs a 3D line cloud by matching 2D segments across views.
• Bayesian probabilistic modeling reconstructs a complete and manifold surface mesh from the 3D line segments.

Main Results:

• The framework successfully reconstructs 3D surfaces using line segments.
• Generated models are abstract, compact, and manifold surface meshes.
• The approach demonstrates validity and low computational cost in experiments.

Conclusions:

• The proposed line segment-based framework offers an effective solution for 3D man-made structure reconstruction.
• It overcomes limitations of existing methods, particularly for challenging datasets.
• The method is computationally efficient and suitable for hardware with low performance requirements.