Efficient Large-Scale Structure From Motion by Fusing Auxiliary Imaging Information.

Area of Science:

• Computer Vision
• Photogrammetry
• 3D Reconstruction

Background:

• Large-scale 3D scene reconstruction benefits from global approaches utilizing auxiliary data.
• Auxiliary information (GPS, IMU, EXIF) is often noisy and not directly usable.
• Existing methods may struggle with noisy data in structure from motion.

Purpose of the Study:

• To develop robust global optimization algorithms for 3D scene reconstruction.
• To leverage noisy auxiliary information for improved structure from motion (SfM).
• To enhance the accuracy, completeness, and scalability of large-scale 3D reconstructions.

Main Methods:

• Introduced two iterative global optimization algorithms for SfM.
• Developed a robust rotation averaging algorithm for noisy epipolar graphs.
• Implemented a robust scene reconstruction algorithm using noisy GPS data for camera initialization.
• Focused on estimated inliers during iterative optimization for convergence.

Main Results:

• The proposed algorithms effectively utilize noisy auxiliary data for SfM.
• Optimization converges well and efficiently by focusing on inliers.
• Evaluated on diverse datasets including UAV, StreetView, and digital cameras.
• Achieved comparable or superior performance to state-of-the-art methods in accuracy and completeness.

Conclusions:

• The novel approach significantly improves large-scale 3D scene reconstruction.
• The method demonstrates enhanced efficiency and scalability for complex image datasets.
• Robust handling of noisy auxiliary information is key to successful global SfM.