Multi-Level Active Prediction of Useful Image Annotations for Recognition