A unifying approach to feature point orientation assignment
(English)Manuscript (preprint) (Other academic)
A general structure of an orientation measure for orientation assignment of 2D image feature points is heuristically motivated. The measure is discretized and general approximation methods are presented. Orientation assignment methods found in the literature are shown to exploit special cases of the general measure, which thereby provides a unifying framework for them. An analytical robustness analysis is conducted, giving a number of desirable robustness properties of the measure components. Following this, a detailed treatment of implementation issues such as gradient sampling and binning is given and based on the desirable properties and other implementation considerations, specific measure components with implementations are suggested. Together, this constitutes a novel feature point orientation assignment method, which we have called RAID. We argue that this method is considerably less expensive than comparable methods in the literature, and by means of a quantitative perturbation analysis, a significantly improved orientation assignment repeatability is demonstrated compared with the available methods found in the literature.
Computer Vision and Robotics (Autonomous Systems)
IdentifiersURN: urn:nbn:se:kth:diva-133475OAI: oai:DiVA.org:kth-133475DiVA: diva2:661848
QS 20132013-11-052013-11-052013-11-05Bibliographically approved