Kinship Verification through Transfer Learning

Siyu Xia, Ming Shao, Yun FuIJCAI 2011 · OpenReview

Abstract

Because of the inevitable impact factors such as pose, expression, lighting and aging on faces, identity verification through faces is still an unsolved problem. Research on biometrics raises an even challenging problem--is it possible to determine the kinship merely based on face images? A critical observation that faces of parents captured while they were young are more alike their children's compared with images captured when they are old has been revealed by genetics studies. This enlightens us the following research. First, a new kinship database named UB KinFace composed of child, young parent and old parent face images is collected from Internet. Second, an extended transfer subspace learning method is proposed aiming at mitigating the enormous divergence of distributions between children and old parents. The key idea is to utilize an intermediate distribution close to both the source and target distributions to bridge them and reduce the divergence. Naturally the young parent set is suitable for this task. Through this learning process, the large gap between distributions can be significantly reduced and kinship verification problem becomes more discriminative. Experimental results show that our hypothesis on the role of young parents is valid and transfer learning is effective to enhance the verification accuracy.

Introduction

Kinship verification—determining if two individuals are biologically related based on facial images—represents a significant challenge in biometrics. Unlike standard facial recognition, kinship verification must contend with the profound effects of aging. A child’s face and their parent’s face, often captured decades apart, belong to vastly different statistical distributions due to changes in skin texture, sag, and bone structure. These factors frequently mask underlying genetic similarities.

Existing kinship databases often feature images taken in controlled environments, which do not reflect the "in the wild" variability of real-world photography. This paper addresses these gaps by introducing a novel dataset that tracks parents across time and proposing a transfer learning algorithm designed to "see through" the aging process to find kinship signals.

The UB KinFace Dataset

To facilitate this research, the authors introduced the UB KinFace (Ver 1.0) benchmark. This dataset consists of 270 images representing 180 individuals organized into 90 groups. Each group is a triplet containing: (1) a child's face, (2) the parent's face when they were young, and (3) the same parent's face when they were old.

Examples from the UB KinFace dataset showing (1) child, (2) young parent, and (3) old parent triplets across different families.

The images were sourced from the internet, featuring celebrities and public figures to ensure a wide range of uncontrolled lighting, poses, and expressions, as shown in the broader dataset samples.

Additional samples from the UB KinFace dataset illustrating the "in the wild" variability in pose, expression, and lighting.

Methodology: Extended Transfer Subspace Learning

The core innovation is an Extended Transfer Subspace Learning (TSL) framework. Traditional transfer learning attempts to map a source distribution (child) directly to a target distribution (old parent). However, the distribution gap caused by aging is often too wide for direct mapping.

The Intermediate Bridge

The proposed method leverages the "young parent" distribution as a mathematical bridge. Genetics suggest that facial resemblance is strongest at similar life stages. By using the young parent as an intermediate link, the algorithm can more effectively transfer resemblance knowledge from the child-young pair to the child-old pair.

Subspace Optimization

The goal is to find an optimal projection matrix $W$ that maps all three groups into a single, low-dimensional discriminative subspace. The algorithm employs Bregman divergence-based regularization to simultaneously minimize the distances between the Child/Young-Parent and Young-Parent/Old-Parent distributions. By combining this with techniques like Discriminative Locality Alignment, the model creates a space where the "kinship signal" is louder than the "aging noise."

Experimental Results and Insights

The authors' hypothesis—that young parents share a higher facial resemblance with their children—was validated using Cumulative Match Characteristic (CMC) curves. Key findings include:

  • Verification Accuracy: The proposed TSL method achieved 60.00% accuracy using Gabor features on specific facial regions, outperforming standard TSL (56.11%).
  • Feature Importance: Focusing on specific facial regions (eyes, nose, and mouth) proved more effective than full-face analysis, as these features are more genetically "fixed."
  • Structural Advantage: Anthropometric "Structure" features generally outperformed raw pixel data, highlighting the importance of facial geometry in kinship.

Conclusion and Future Work

This research demonstrates that leveraging intermediate life stages through transfer learning significantly improves kinship verification. While the current UB KinFace dataset is a valuable starting point, future work will focus on expanding the dataset size (Version 2.0), exploring illumination-invariant features, and comparing machine performance against human intuition in kinship detection.

Summary updated

Abstract

Because of the inevitable impact factors such as pose, expression, lighting and aging on faces, identity verification through faces is still an unsolved problem. Research on biometrics raises an even challenging problem--is it possible to determine the kinship merely based on face images? A critical observation that faces of parents captured while they were young are more alike their children's compared with images captured when they are old has been revealed by genetics studies. This enlightens us the following research. First, a new kinship database named UB KinFace composed of child, young parent and old parent face images is collected from Internet. Second, an extended transfer subspace learning method is proposed aiming at mitigating the enormous divergence of distributions between children and old parents. The key idea is to utilize an intermediate distribution close to both the source and target distributions to bridge them and reduce the divergence. Naturally the young parent set is suitable for this task. Through this learning process, the large gap between distributions can be significantly reduced and kinship verification problem becomesmore discriminative. Experimental results show that our hypothesis on the role of young parents is valid and transfer learning is effective to enhance the verification accuracy.
BibTeX
@inproceedings{xia2011kinship,
  author = {Siyu Xia and Ming Shao and Yun Fu},
  title = {{Kinship Verification through Transfer Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2011},
  url = {https://openreview.net/forum?id=H1ZcXrMd-H},
}
Figure