Feature-based Object Recognition - a case study for car model detection

Recognition of untextured as well as shiny objects introduces problems to most of the present interest point detectors and descriptors. Those problems arise from changes in brightness intensities around potential interest points, caused by reflections on such an object's surface. Since the selection of the best performing recognition algorithm heavily depends on the underlying task, an evaluation of some algorithms tested especially for the respective data collection and typical query types has to be taken into account. If the task is for example the model detection of a car, most of the existing algorithm evaluations are misleading since they use data sets with mostly textured and matt objects. The present paper introduces a data set of car images which stands for untextured, shiny objects and evaluates different algorithms namely SIFT, ASIFT, SURF, ORB, and BRISK on that data set. Findings are that most algorithms have a low effectiveness recognizing such objects. This mainly depends on the kind of interest point detector and is not so much dependent on the kind of descriptor. Detectors preferring points on edges and corners are more reliable than those preferring interest points on blob-like structures.