Mir: Multi-Task Intra-Representation Learning for Face Classification

摘要

Despite significant recent advances in the field of face recognition, implementing face recognition efficiently on a small scale datasets and how to learn a better intra-representation of the face present serious challenges. In this paper we present a new method called MIR, that directly learns similarity feature between face and face and uses characteristic as a new representation to do classification. Our method have a better performance with the characteristic than original features in different datasets(LFW, Youtube Faces(YTF) and MegaFace challenge). Our method describe the effects of face classifications after learning similarity and demonstrate that multitasking get a better accuracy than single task learning tasks.