Mapping image data into the embedding space where objects of the same class or label have a short distance in-between and objects of different classes have long margins, has been an essential task for many computer vision applications. However, current approaches struggle to map image data into a proper embedding space due to the difficulty of constructing discriminative features from among numerous features from the original data. Existing approaches include finding effective loss and new sampling methods, which do not consider improving the embedding space by selecting fine features extracted by the network. In this work, we proposed a new attention approach by exploiting the variance of features. The method can improve the performance of the current metric learning. Our approach consists of a variance estimation module(VEM) and fusion stage for applying channel-wise attention on extracted features. It is easy to implement and fast for training. Unlike other traditional second-order based methods, the variance estimation module does not embed second-order calculation in the network itself, and cost no large extra computation time in the evaluation stage. The experiment shows promising performance while compared with current SOTA approaches on multiple metric learning benchmark datasets such as CUB200-2011, CARS196, In-shop Clothes. Contribution-We design a new attention module by using estimation of variance in the features and achieve SOTA results in several benchmarks with almost no extra time cost in the test stage.