Panoramic Horizon Recognition
David Rawlinson, Ray A. Jarvis · 2008
This paper describes a method in which SIFT features on and around the horizon are used for image classification (matching or recognition). Ordinarily, transformational geometry relating two images is recovered from a set of point correspondances – but correspondance formation may be improved and accelerated using geometric constraints. The proposed solution to this chicken/egg problem is to first search for an approximate geometric relationship between images using only the scale, orientation and position of SIFT features; if found, the recovered transformational geometry is used to filter feature matching. Generous feature position tolerance simplifies the search to 1 dimension in unwarped panoramic images.