Learning object recognition models from images