Scale-invariant features and polar descriptors in omnidirectional imaging

Area of Science:

• Computer Vision
• Geometric Deep Learning
• Robotics

Background:

• Omnidirectional images present unique geometric challenges for feature detection.
• Existing scale-invariant feature transform (SIFT) methods struggle with the non-Euclidean geometry of spherical images.

Purpose of the Study:

• To develop a novel framework for computing scale-invariant features in omnidirectional images.
• To improve feature detection and matching accuracy for spherical visual data.
• To enable feature matching between omnidirectional images with varying geometric properties.

Main Methods:

• Utilized Riemannian geometry to define differential operators adapted to omnidirectional imaging systems.
• Developed a scale-space analysis preserving the geometric integrity of visual information.
• Introduced a new descriptor based on log-polar transformations, optimized for non-uniform sampling.
• Proposed a rotation-invariant matching technique to reduce computational complexity.

Main Results:

• The proposed method demonstrates superior detection and matching performance over standard SIFT and spherical SIFT on unwrapped omnidirectional images.
• The framework successfully maps features onto the sphere, maintaining geometric consistency.
• Achieved effective feature matching between images captured with different camera geometries.

Conclusions:

• The novel Riemannian geometry-based approach offers a robust solution for scale-invariant feature computation in omnidirectional imaging.
• The adapted descriptor and rotation-invariant matching significantly enhance performance and efficiency.
• This framework advances the capabilities of computer vision systems dealing with spherical imagery.