MetaAge: Meta-Learning Personalized Age Estimators.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Individual aging varies significantly, necessitating personalized age estimation methods.
• Current personalized age estimation techniques require large datasets with identity labels and long-term aging patterns, which are often unavailable.
• Existing methods struggle with data scarcity for individual aging trajectories.

Purpose of the Study:

• To develop a personalized age estimation method that does not require extensive identity-specific data.
• To propose a meta-learning framework, MetaAge, for learning personalized age estimators.
• To overcome the limitations of data requirements in current personalized age estimation.

Main Methods:

• Proposed MetaAge, a meta-learning method that learns a mapping from identity features to age estimator parameters.
• Introduced a personalized estimator meta-learner that takes identity features as input and outputs customized estimator parameters.
• Leveraged existing large-scale age datasets without additional annotations by learning transferable meta-knowledge.

Main Results:

• MetaAge significantly enhances the performance of existing personalized age estimation methods.
• The proposed method outperforms state-of-the-art approaches on benchmark datasets (MORPH II, ChaLearn LAP 2015, ChaLearn LAP 2016).
• Demonstrated effective personalization of age estimation even with limited individual data.

Conclusions:

• MetaAge offers a robust solution for personalized age estimation by learning from identity information.
• The meta-learning approach effectively addresses the challenge of data scarcity in personalized aging studies.
• MetaAge represents a significant advancement in accurate and personalized age estimation technology.