Li Wanhua ; Zhang Yingqiang ; Lv Kangchen ; Lu Jiwen ; Feng Jianjiang ; Zhou Jie

Abstract

In this paper, we propose a graph-based kinship reasoning (GKR) network forkinship verification, which aims to effectively perform relational reasoning onthe extracted features of an image pair. Unlike most existing methods whichmainly focus on how to learn discriminative features, our method considers howto compare and fuse the extracted feature pair to reason about the kinrelations. The proposed GKR constructs a star graph called kinship relationalgraph where each peripheral node represents the information comparison in onefeature dimension and the central node is used as a bridge for informationcommunication among peripheral nodes. Then the GKR performs relationalreasoning on this graph with recursive message passing. Extensive experimentalresults on the KinFaceW-I and KinFaceW-II datasets show that the proposed GKRoutperforms the state-of-the-art methods.