Distributed Very Large Scale Bundle Adjustment by Global Camera Consensus.

Area of Science:

• Computer Vision
• Photogrammetry
• Computational Geometry

Background:

• Conventional Structure-from-Motion (SfM) methods face scalability challenges due to monolithic global bundle adjustment.
• The computational complexity of all-in-one optimization limits SfM applications on very large datasets.

Purpose of the Study:

• To develop a distributed optimization framework for large-scale SfM, addressing the limitations of traditional methods.
• To enhance the efficiency and accuracy of SfM computation for massive image collections.

Main Methods:

• A distributed formulation based on the Alternating Direction Method of Multipliers (ADMM) is derived for global camera consensus.
• Convergence guarantees for the distributed optimization are analyzed, incorporating over-relaxation and self-adaptation for improved rates.
• The camera-point visibility graph is partitioned to minimize communication overhead in distributed computing.

Main Results:

• The proposed distributed SfM method demonstrates superior performance compared to state-of-the-art techniques.
• Experiments on public and proprietary large-scale aerial datasets validate the efficiency and accuracy gains.
• The approach effectively handles the computational demands of very large-scale SfM.

Conclusions:

• The distributed ADMM-based approach offers a scalable and efficient solution for large-scale Structure-from-Motion.
• This work significantly advances the practical application of SfM in areas requiring massive data processing.
• The proposed method provides a robust alternative to conventional global bundle adjustment for complex SfM problems.