Inheritable Fisher vector feature for kinship verification

Abstract

An innovative inheritable Fisher vector feature (IFVF) method is presented in this paper for kinship verification. Specifically, Fisher vector is first derived for each image by aggregating the densely sampled SIFT features in the opponent color space. Second, a new inheritable transformation, which maximizes the similarity between kinship images while minimizes that between non-kinship images for each image pair simultaneously, is learned based on the Fisher vectors. As a result, the IFVF is derived by applying the inheritable transformation on the Fisher vector for each image. Finally, a novel fractional power cosine similarity measure, which shows its theoretical roots in the Bayes decision rule for minimum error, is proposed for kinship verification. Experimental results on two representative kinship data sets, namely the KinFaceW-I and the KinFaceW-II data sets, show the feasibility of the proposed method.