City-Scale Localization for Cameras with Known Vertical Direction.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Localizing novel images within large-scale 3D models is crucial for applications like augmented reality and autonomous navigation.
• The significant number of outliers in correspondence matching poses a major challenge for traditional camera pose estimation methods.

Purpose of the Study:

• To develop a robust camera pose estimation technique for large 3D models, effectively handling a high percentage of outlier correspondences.
• To leverage gravitational sensor data to reduce the search space and improve localization accuracy.

Main Methods:

• Utilizing gravitational sensor data from modern cameras and phones to constrain the search space for camera pose.
• Extending recent outlier rejection techniques with accurate approximations and fast polynomial solvers for robust correspondence matching.
• Applying these methods to camera pose estimation in large-scale 3D environments.

Main Results:

• Demonstrated reliable camera pose estimation even with more than 99% outlier correspondences.
• Successfully localized images within city-scale 3D models comprising millions of 3D points.
• Achieved accurate camera pose estimation in challenging scenarios with extreme outlier rates.

Conclusions:

• The proposed method offers a robust and efficient solution for camera pose estimation in large-scale 3D models.
• The integration of gravitational sensing and advanced outlier handling significantly enhances localization reliability.
• This approach paves the way for more accurate and dependable navigation and mapping systems.