Virtually augmenting hundreds of real pictures: An approach based on learning, retrieval, and tracking