Toward kinship verification using visual attributes

In this paper, we consider the kinship verification problem through face images. Genetics studies show resemblant regions on faces among immediate family members are mostly concentrated on eyes, nose, mouth, etc. This motivates us the following research. First, we construct low-level features based on the hierarchical local regions. Second, an attribute based method is proposed towards meaningful middle-level representations for face images. Specially, we use LIFT algorithm to extract binary attributes, e.g., “male”, “has glasses”. Simultaneously, relative attributes, e.g., “bigger” eyes, are learned through comparing with reference images. A SVM classifier is trained based on the combination of these two attributes. Experimental results demonstrate the effectiveness of our method. Finally, we show the heredity of attributes in terms of four kin relations.