Are Minimal Radial Distortion Solvers Really Necessary for Relative Pose Estimation?

Area of Science:

• Computer Vision
• Robotics
• Photogrammetry

Background:

• Estimating relative camera pose is crucial for Structure-from-Motion and other applications.
• Standard methods use minimal solvers within RANSAC loops, but often neglect radial distortion.
• Ignoring radial distortion significantly degrades pose estimation accuracy.

Purpose of the Study:

• To evaluate simpler alternatives to complex minimal solvers for radial distortion in relative pose estimation.
• To compare a sampling-based undistortion approach with a learning-based method.
• To determine when simpler methods outperform traditional complex solvers.

Main Methods:

• Compared efficient pinhole solvers with sampled radial undistortion parameters.
• Evaluated a neural network approach for estimating distortion parameters.
• Conducted extensive experiments on diverse datasets and camera setups.

Main Results:

• Complex minimal radial distortion solvers are not essential in practice.
• Simpler methods achieve comparable results to complex solvers.
• Performance depends on specific dataset and camera configurations.

Conclusions:

• Effective relative pose estimation under radial distortion can be achieved without complex minimal solvers.
• Simple radial undistortion parameter sampling is often preferable to learning-based calibration.
• The study provides a benchmark and code for further research.