A New Similarity Space Tailored for Supervised Deep Metric Learning

We propose a novel deep metric learning method. Differently from many works in this area, we define a novel latent space obtained through an autoencoder. The newspace, namely S-space, is divided into different regions describing positions where pairs of objects are similar/dissimilar. We locate makers to identify these regions and estimate the similarities between objects through a kernel-based Cauchy distribution to measure the markers' distance and the new data representation. In our approach, we simultaneously estimate themarkers' position in the S-space and represent the objects in the same space. Moreover, we propose a newregularization function to prevent similar markers from collapsing altogether. Our method emphasizes the group property (separability) while preserving instance representativity. We present evidence that our proposal can represent complex spaces, for instance, when groups of similar objects are located in disjoint regions. We compare our proposal to nine different distance metric learning approaches (four of them are based on deep learning) on 28 real-world heterogeneous datasets. According to the four quantitative metrics used, our method overcomes all of the nine strategies from the literature.

Automated summary

We propose a novel deep metric learning method using an autoencoder to define a new latent space called S-space. We locate markers in S-space to identify similar and dissimilar objects and estimate their similarities using a kernel-based Cauchy distribution. Our approach simultaneously estimates the markers' positions and represents the objects in the same space while preventing collapsing of similar markers. We demonstrate the effectiveness of our method on various datasets.