In experiments, the proposed algorithm significant outperforms both simple random sampling as well as other types of stratied sampling, with an average reduction of about 20% in labeling effort to reach the same confidence and interval-bounds on precision.