Metric Learning with Quadruplets on Non-Hierarchical Labeled Datasets

Abstract

Metric learning is a frequently utilized method that can be applied to many different computer vision problems and; thus, facilitates these problems. In this paper, the proposed method for manipulating the feature space with the help of quadruplets, which is one of the metric learning methods, is explained with details, and the results obtained by using it are shown. Many methods used in the field of metric learning, have been developed on the Siamese and triplet structures and satisfactory results have been obtained by them. Similarly, the quadruplet structure is still being studied and different approaches are proposed in the literature. How to select four elements for generating a quadruplet is not standardized at present. However, by looking at the mining methods of Siamese and triplet samples, it can be concluded that each sample in the dataset that is trained with quadruplet samples, must have at least two hierarchical labels. This makes it possible to train the quadruplet structure only on certain datasets that have the hierarchical labels. In this paper, an unsupervised method is proposed to use the quadruplet structure in datasets that do not have a hierarchical label structure. Besides, the performances of three different quadruplet mining methods are compared and an ablation study is conducted by discussing the advantages and disadvantages of the proposed method.