Affine Correspondences Between Multi-Camera Systems for Relative Pose Estimation.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Multi-camera relative pose estimation is crucial for robotics and computer vision.
• Existing methods often suffer from high computational complexity or require numerous point correspondences (PCs).
• These limitations hinder efficient and accurate pose estimation, especially with robust estimators like RANSAC.

Purpose of the Study:

• To develop a novel, efficient, and accurate method for multi-camera relative pose estimation.
• To address the limitations of existing solvers, particularly in scenarios with limited correspondences or unknown scale.
• To provide a unified framework for generating 6DOF (Degrees of Freedom) solvers.

Main Methods:

• Utilizing two affine correspondences (ACs) for 6DOF relative pose estimation.
• Developing a special parameterization to exploit geometric constraints between ACs and multi-camera systems.
• Proposing 7DOF solvers using three ACs for scale-unknown problems.
• Introducing minimal solvers leveraging inertial measurement units (IMUs) for constrained problems (5DOF and 4DOF).

Main Results:

• The proposed method enables feasible minimal solutions for 6DOF relative pose estimation using ACs.
• Achieved efficient and accurate 7DOF solvers for scale-ambiguous scenarios.
• Demonstrated improved efficiency and accuracy of proposed solvers over state-of-the-art methods in both virtual and real-world experiments.
• Successfully introduced constrained solvers (5DOF, 4DOF) using IMUs.

Conclusions:

• The novel approach using affine correspondences significantly advances multi-camera relative pose estimation.
• The proposed solvers offer superior efficiency and accuracy, overcoming limitations of prior methods.
• The unified framework provides versatile solutions for various practical scenarios, including scale-ambiguous and IMU-aided estimations.