Exhaustive linearization for robust camera pose and focal length estimation.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Accurate camera pose and focal length estimation is crucial for 3D scene understanding and augmented reality.
• Existing methods like EPnP (Efficient PnP) offer efficient solutions for calibrated cameras but struggle with uncalibrated scenarios or noisy data.

Purpose of the Study:

• To develop a novel, accurate, and efficient method for estimating both camera pose and focal length simultaneously from 3D-to-2D point correspondences.
• To address the limitations of existing methods when dealing with uncalibrated cameras and significant noise.

Main Methods:

• Introduced novel 'exhaustive linearization' and 'exhaustive relinearization' techniques to overcome limitations of traditional methods when focal length is unknown.
• Developed a closed-form solution inspired by the EPnP algorithm, adapted for simultaneous pose and focal length estimation.
• Systematically explored the solution space to handle noise effectively.

Main Results:

• The proposed method achieves precise focal length estimation.
• Retrieved camera pose accuracy is comparable to the EPnP algorithm, which assumes a calibrated camera.
• Demonstrated superior accuracy and speed compared to existing closed-form and iterative solutions, especially under noisy conditions.

Conclusions:

• The novel approach provides a robust and efficient solution for simultaneous camera pose and focal length estimation.
• This method significantly advances uncalibrated camera pose estimation, particularly in challenging, noisy environments.
• The technique offers a valuable tool for applications requiring accurate camera parameter recovery without prior calibration.