The proposed method dramatically reduces the computational cost by avoiding the costive samplewise affinity estimation, without sacrificing accuracy.