Ordinal space projection learning via neighbor classes representation

Abstract

Ordinal metric learning (OML) is an important research branch of metric learning and has attracted increasing attention due to its wide applications where there typically exists a global order (i.e., ordinal) relationship among the data classes, such as human age estimation and head pose recognition, etc. Although several works have been proposed to perform OML, they suffer from two critical drawbacks: 1) only the local ordinal relationships among classes rather than the global one are considered, and 2) the very time-consuming semi-definite programming is commonly involved. To avoid these shortcomings, in this paper we propose a novel OML approach, named ordinal space projection learning (OSPL) by representing one class with its neighbor classes whose representation weights are proportional to their class distance (i.e., absolute label difference), in which the neighboring ordinal class similarities are taken into account in their reconstruction. Then, we present an alternating optimization algorithm to solve the proposed method to learn an ordinal distance space. Finally, extensive experiments on ordinal estimation tasks demonstrate effectiveness and superiority of the proposed method.