I. Introduction
Deep metric learning (DML) has been actively studied recently due to its widespread applications, such as image retrieval and classification [1], [2], [3], visual tracking [4], person re-identification [5], [6], [7], and face verification [8], [9], [10], [11]. The core idea of DML is to seek a reliable embedding space by virtue of nonlinear deep neural networks, where samples from the same class are encouraged to be closer than those from different classes.