Adaptive feature selection for kinship verification

Kinship verification from facial images is a challenging task in computer vision. The majority of recent verification algorithms concatenate all features of patches in facial image to build the final feature representation, which implicitly takes every facial part into account for kinship verification. However, it is questionable by considering all face regions since certain facial parts such as the eyes of a child look like the other parent or someone else, which could decrease the verification performance of an algorithm. Motivated by this observation, we attempt to automatically single out the discriminative patches of face and discard interferential ones. Specifically, each weak classifier of AdaBoost is adaptively trained to select a feature composed of some crucial feature patches and various features can be obtained by different weak classifiers. In the stage of verification, a strong classifier constructed by these weak classifiers ensembling is utilized to merge the contributions of selected distinct face areas. Experimental results on KinfaceW-I and KinfaceW-II datasets demonstrate the efficacy of the proposed method compared with state-of-the-art kinship verification approaches.