Recognizing Multigenerational Families From Videos in the Wild

In the field of computer vision, current relative learning research has predominantly concentrated on identifying parent-child relationships from pairs of static facial images, neglecting the discriminative information inherent in relatives involving multiple subjects and disregarding the common noise in unconstrained settings. To tackle these limitations, we present a novel task of video-based multigenerational family recognition, aiming to recognize multigenerational families from videos captured in unconstrained environments. We propose a Support Vector Data Description (SVDD)-based family-shared multi-metric learning (SFM2L) method, where only purified samples are subjected to multi-metric learning to derive both family-shared and family-specific distance metrics. To further improve the recognition performance, we introduce a multi-view method named MSFM2L, which effectively integrates deep and shallow features. In addition, we have constructed a new video dataset consisting of 90 multigenerational families. Extensive experiments on both the newly collected dataset and the well-established KinFaceW kinship face dataset clearly demonstrate the superior performance of our proposed methods compared to existing metric learning approaches.