RAgE: Robust Age Estimation Through Subject Anchoring With Consistency Regularisation.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Facial age estimation systems perform well under controlled conditions.
• Domain shifts in real-world data significantly degrade the accuracy of existing age estimation algorithms.
• Robustness to variations in pose, illumination, and expression is a key challenge.

Purpose of the Study:

• To develop a novel method, RAgE, for robust facial age estimation.
• To leverage unlabelled data to improve age estimation accuracy and reduce uncertainty.
• To enhance system resilience against domain shifts and confounding factors.

Main Methods:

• Proposed a similarity-preserving pseudo-labelling algorithm for unlabelled data.
• Introduced a novel consistency regularisation term for output invariance.
• Utilized a subject anchoring strategy to group images of the same individual.
• Developed a noise-tolerant regularisation term to mitigate confirmation bias from pseudo-labels.

Main Results:

• RAgE demonstrated substantial improvements over state-of-the-art methods on benchmark datasets.
• The method showed significant robustness to variations in head pose, illumination, and facial expression.
• Experiments confirmed the effectiveness of the proposed pseudo-labelling and consistency regularisation techniques.

Conclusions:

• RAgE offers a robust solution for facial age estimation in the presence of domain shifts.
• Leveraging unlabelled data through subject anchoring and consistency regularisation is effective.
• The proposed method significantly advances the state-of-the-art in unconstrained facial age estimation.