Minimal Case Relative Pose Computation Using Ray-Point-Ray Features.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Corners are fundamental features for 2D-2D point correspondence in relative pose estimation.
• Ray-point-ray (RPR) structures, formed by two 3D rays sharing a common origin, offer unique geometric properties beyond simple keypoints.

Purpose of the Study:

• To investigate the additional constraints provided by Ray-point-ray (RPR) features in two-view geometry.
• To determine the minimal number of RPR correspondences required for relative pose computation.

Main Methods:

• Analysis of reprojection properties of RPR structures, including orientation and inscribed angle.
• Derivation of constraints on relative orientation using known inscribed angles of RPR features.
• Investigation of minimal cases, including 90-degree RPR structures, arbitrary known angles, directional correspondence, and planar motion.

Main Results:

• Relative pose can be solved using as few as three RPR correspondences by leveraging inscribed angle constraints.
• Planar motion scenarios are solvable with a single RPR correspondence.
• A robust RPR feature extraction technique is presented.

Conclusions:

• RPR features offer a powerful new approach to relative pose estimation, reducing the number of required correspondences.
• The method shows high potential for applications in man-made environments where 90-degree RPR structures are prevalent.