Keypoint detection and matching on high resolution spherical images

Abstract

In this paper we address one of the fundamental topics in Computer Vision: feature detection and matching. We propose a new approach for detecting interest points on high-resolution spherical images. To achieve invariance to scale changes, points are detected across scale-space representations of images. Instead of using domain transformations or special projection approximations, we build the scale-space by solving the heat diffusion equation directly on the sphere surface. Interest points are defined as an extension of the classical Harris corner detector to the sphere. Our technique is validated on a set of real world spherical images. We perform a quantitative evaluation of our approach and obtain results that demonstrate its suitability to higher level applications, such as structure from motion and camera registration. © 2013 Springer-Verlag.