Deep metric learning via subtype fuzzy clustering

Highlights：

• We exploit clustering structure in each class to achieve discriminant metric learning. The refined subtype clusters, which can be viewed as magnification by a microscope, are expected to enhance the separability between different classes. These subtype clusters are then used to construct the discriminant and down-scaled triplet loss.

• We define the concept of clustering degree for each sample pair. The clustering degree is inversely proportional to the distance between two samples in a positive pair. By using the clustered data to construct triplet loss function, it is equivalent to mine easy sample pairs to build positive pairs. In contrary, traditional triplet loss-based methods simply focus on the hard or semi-hard pairs mining.

• The clustering module can be transferred to neighborhood selection problem. We propose an online sampling method for choosing positive pairs with high clustering degrees. Therefore, this online sampling approach helps us to select desirable positive pairs efficiently in the training procedure.