Multi-attribute and Multi-label Deep Metric Learning via Pair-based and Proxy-based Losses

Shozo Saeki, Minoru Kawahara, Hirohisa Aman

Abstract: The main goal of deep metric learning (DML) is to train a neural network that can accurately map data onto an embedding space, placing similar data close together and dissimilar data far apart. Conventional DML postulates that each data has a single positive class, and data similarities rely on whether they belong to the same class. However, similarity could depend not only on a single perspective but also on multiple perspectives. To address this, we introduce multi-attribute and multi-label deep metric learning (MADML), which postulates that each has have multi-positive classes. We propose multi-attribute and multi-label triplet (MAT) loss and multi-attribute and multi-label soft triple (MAST) loss to solve MADML problems. Finally, we demonstrate the performance of ranking retrievals. As a result, we find that MAST loss has a higher performance and can retrieve rankings. This study has two main contributions. (1) MAT and MAST losses can be applied to MADML problems without changing anything other than the loss function. (2) The networks trained using MAT and MAST losses can retrieve the rankings of any rating.