Q-learning of sequential attention for visual object recognition from informative local descriptors