Knowledge-Guided Symbolic Regression for Interpretable Camera Calibration.

Area of Science:

• Computer Vision
• Optical Engineering
• Machine Learning

Background:

• Accurate camera calibration is crucial for applications like autonomous navigation and robotics.
• Existing models (pinhole, radial-tangential) and neural methods have limitations in handling complex distortions, data requirements, and computational cost.

Purpose of the Study:

• To develop a flexible, interpretable, and resource-efficient symbolic model discovery framework for camera calibration.
• To identify lens-specific projection and distortion models that accurately represent optical deviations.

Main Methods:

• Utilized symbolic regression and genetic programming (GP) guided by physical knowledge.
• Incorporated established distortion models (Brown-Conrady, Mei-Rives, Kannala-Brandt, double-sphere) as modular components.
• Embedded models into the symbolic search to constrain the solution space for efficient parameter fitting and model selection.

Main Results:

• The framework successfully identified calibration models for various lens types (fisheye, omnidirectional, catadioptric, traditional).
• Achieved comparable or superior results to existing calibration techniques.
• Demonstrated robustness against overfitting and efficiency in parameter fitting.

Conclusions:

• The proposed symbolic model discovery framework offers a flexible and interpretable alternative for camera calibration.
• It is particularly suitable for scenarios with scarce calibration data or limited computational resources.
• The method provides accurate and efficient lens distortion modeling.