Relative Pose Estimation for Light Field Cameras Based on LF-Point-LF-Point Correspondence Model

Area of Science:

• Computer Vision
• Robotics
• Optical Engineering

Background:

• Light field (LF) cameras capture 3D scene information using micro-lens arrays (MLAs).
• Accurate relative pose estimation is crucial for applications like 3D reconstruction and navigation.
• Existing methods often struggle with depth estimation inaccuracies due to the small baseline in LF cameras.

Purpose of the Study:

• To propose a novel relative pose estimation algorithm for MLA-based LF cameras.
• To enhance accuracy by avoiding depth recovery and utilizing disparity estimation.
• To provide a robust solution for challenging LF camera scenarios.

Main Methods:

• Establishing an LF-point-LF-point correspondence model using matched LF-point pairs.
• Employing the correspondence model for relative camera pose estimation via linear and non-linear optimization.
• Utilizing estimated disparities instead of recovered depths to circumvent depth inaccuracy issues.

Main Results:

• The proposed algorithm effectively estimates relative camera pose for LF cameras.
• Experimental results on simulated and real data validate the algorithm's performance.
• The method demonstrates superior effectiveness compared to classical and state-of-the-art algorithms.

Conclusions:

• The developed algorithm offers a more accurate and robust approach to relative pose estimation in LF imaging.
• By leveraging disparity estimation, the method overcomes limitations of depth-based techniques.
• This work contributes a significant advancement for LF camera applications requiring precise spatial understanding.