IV.EXPERIMENTSIn this section, we present experimental results evaluating the proposed approach. We would like to find an answer for two questions: (a) whether the network is able to learn to better distinguish between similar and dissimilar pairs utilising this kind of training data; (b) if our network is learnable, do features extracted from it generalize to other datasets? In each experiments, we take the activations of fc7 layer of the network as feature vectors.