Evaluation of periocular features for kinship verification in the wild

Kinship verification is receiving increasing attention among computer vision researchers due to interesting applications ranging from family album management to searching missing family members. Existing approaches have focused on using face images to decode kinship information. In contrast, this paper explores the effectiveness of periocular region in verifying kinship from images captured in the wild. Further, we also propose a block-based neighborhood repulsed metric learning (BNRML) framework, an extension of NRML, to yield more discriminative power. The proposed method learns multiple local distance metrics from different blocks of the images represented by local ternary patterns. Moreover, to contemplate diversity in discrimination power of different blocks, weighted score-level fusion scheme is used to obtain a similarity score of image pair. Extensive experiments on KinFaceW-I and KinFaceW-II datasets demonstrated the potential of periocular features for kinship verification. Furthermore, the fusion of periocular and face traits under BNRML framework provided highly competitive results as compared to state-of-the-art methods.