Globally-Optimal Inlier Set Maximisation for Camera Pose and Correspondence Estimation.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Estimating camera pose from images is crucial for applications like augmented reality and autonomous navigation.
• Current methods often struggle with noisy data and require good initial guesses for camera pose and 2D-3D correspondences.

Purpose of the Study:

• To develop a robust and globally optimal method for simultaneously estimating camera pose and 2D-3D correspondences from a single image and a 3D point set.
• To overcome limitations of existing local optimization techniques and restrictive assumptions.

Main Methods:

• A branch-and-bound approach is employed to search the 6D space of camera poses, ensuring global optimality.
• Novel upper and lower bounds derived from SE(3) geometry are used to prune the search space.
• Local optimization is integrated to accelerate convergence.

Main Results:

• The proposed method reliably finds the global optimum for camera pose and correspondence estimation.
• It outperforms existing approaches on challenging synthetic and real-world datasets.
• A GPU implementation significantly reduces computation time.

Conclusions:

• The branch-and-bound inlier set maximization approach offers a robust and globally optimal solution for camera pose estimation.
• This method eliminates the need for good pose initialization and restrictive assumptions.
• The approach demonstrates superior performance and efficiency in challenging computer vision scenarios.