Appearance Based Salient Point Detection with Intrinsic Scale-frequency Descriptor

Recent object recognition methods propose to represent objects by collections of local appearance descriptors in several interest points. For recognition, this representation is matched to image data. Interest points (candidates for matching) are usually selected from images in a purely bottom-up manner. However, in many situations, there is a limited number of objects to search for, and some information about object characteristics should be employed in the detection of salient points, to reduce the number of potential candidates. In this paper we propose a methodology for the selection of candidates with prior information of the object local appearance. Points are represented by a rotation and scale invariant descriptor, composed by the response of filters derived from Gabor functions, denoted as “intrinsic scale/frequency descriptor”. When compared to classical salient point detectors, like extremal points of laplacian operators at several scales, the proposed methodology is able to reduce the amount of canditates for matching by more than 60%. Since matching is a costly operation, this strategy will improve the efficiency of object recognition methods.