Compared with previous metric learning methods by using low-rank/online/stochastic strategies, that still encounter

We thank the reviewers for their constructive comments and suggestions. We respond to each point individually. R1: More description to highlight the unique contribution. We will add the missing references in "Related Work" and explain the difference with them in the revision. ML 2015] uses the mini-batch strategy to optimize the metric matrix in the positive semidefinite cone.