Learning a Multiple Kernel Similarity Metric for kinship verification

Abstract

Kinship Verification (KV) has recently caught much attention in the computer vision community due to its potential applications ranging from missing children search to social media analysis. Most of the related work focuses either on developing hand-crafted feature representations to describe the faces or on learning the Mahalanobis distance metric to measure the similarity between facial images. Instead, in this paper, we propose a novel Multiple Kernel Similarity Metric (MKSM), in which, different from the Mahalanobis metric, the similarity computation is essentially based on an implicit nonlinear feature transformation. The overall MKSM is a weighted combination of basic similarities and therefore possesses the capacity for feature fusion. The basic similarities are derived from base kernels and local features, and the weights are obtained by solving a constrained linear programming (LP) problem that originates from a Large margin (LM) criterion. Particularly, the LM criterion not only guarantees the generalization on unseen samples when the training set is small, but also leads to sparsity in the weight vector which in turn boosts the efficiency at the prediction stage. Extensive experiments on four publicly available datasets demonstrate the effectiveness of the proposed method.