Object with important gray-level variations tracked by using SIFT-like detectors. A first step consists in SIFT descriptors learning on several views of the object. After that step recognition is achieved on live video acquisition.