Parallel Structure from Motion for Sparse Point Cloud Generation in Large-Scale Scenes

Area of Science:

• Computer Vision
• 3D Reconstruction
• Geospatial Analysis

Background:

• Scene reconstruction, vital for smart cities and surveying, relies on Structure from Motion (SFM) for 3D modeling.
• Traditional SFM struggles with large-scale scenes due to computational limits and time-consuming image matching/geometric filtering.

Purpose of the Study:

• To develop a novel divide-and-conquer framework for efficient and accurate distributed Structure from Motion (SFM) in large-scale scenes.
• To overcome the limitations of single compute nodes in reconstructing complex environments.

Main Methods:

• Utilizes Global Navigation Satellite System (GNSS) information to define GNSS neighborhoods, significantly reducing image matching scope.
• Employs a clustering algorithm to divide the initial camera graph into subgraphs for parallel local SFM processing.
• Integrates and optimizes local camera poses from multiple computing nodes for global registration.

Main Results:

• The proposed framework achieves robust image matching by leveraging GNSS neighbors.
• Distributed processing of local SFM on subgraphs enables efficient registration of cameras in large-scale scenes.
• Experimental results demonstrate accurate and efficient large-scale scene reconstruction.

Conclusions:

• The divide-and-conquer framework effectively addresses the scalability challenges in Structure from Motion.
• The integration of GNSS data and distributed computing offers a robust solution for 3D scene reconstruction.
• This approach enhances the feasibility of large-scale 3D modeling for applications in smart cities and surveying.