A Unified Approach to Kinship Verification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Kin verification, the task of identifying familial relationships from images, often suffers from limited training data.
• Existing methods struggle with small datasets and inherent class imbalance in kin verification datasets.

Purpose of the Study:

• To develop a deep learning approach for kin verification that effectively utilizes small training sets.
• To address overfitting and data imbalance issues common in kin verification tasks.

Main Methods:

• A unified multi-task learning scheme is proposed to jointly learn all kinship classes.
• A novel embedding fusion technique is introduced to prevent overfitting.
• An adaptive sampling scheme is derived to manage imbalanced training data.

Main Results:

• The proposed method demonstrates superior performance on benchmark kin verification datasets (Families In the Wild, FG2018, FG2020).
• Ablation studies confirm the effectiveness of the individual components of the proposed approach.
• The unified multi-task learning and embedding fusion significantly improve utilization of small training sets.

Conclusions:

• The proposed deep learning approach offers a robust solution for kin verification, particularly in low-data regimes.
• The novel techniques for embedding fusion and adaptive sampling enhance model generalization and performance.
• This work sets a new state-of-the-art for kin verification accuracy on established datasets.