Enhancing Facial Data Diversity with Style-based Face Aging

Markos Georgopoulos, James Oldfield, Mihalis A. Nicolaou, Yannis Panagakis, Maja Pantic

Abstract: A significant limiting factor in training fair classifiers relates to the presence of dataset bias. In particular, face datasets are typically biased in terms of attributes such as gender, age, and race. If not mitigated, bias leads to algorithms that exhibit unfair behaviour towards such groups. In this work, we address the problem of increasing the diversity of face datasets with respect to age. Concretely, we propose a novel, generative style-based architecture for data augmentation that captures fine-grained aging patterns by conditioning on multi-resolution age-discriminative representations. By evaluating on several age-annotated datasets in both single- and cross-database experiments, we show that the proposed method outperforms state-of-the-art algorithms for age transfer, especially in the case of age groups that lie in the tails of the label distribution. We further show significantly increased diversity in the augmented datasets, outperforming all compared methods according to established metrics.

Introduction

Artificial Intelligence is only as good as the data it's trained on. In facial analysis, most widely used datasets are heavily biased toward younger individuals; for example, 87% of the popular FG-NET dataset consists of people under 30. This "age bias" leads to algorithms that perform poorly or unfairly on older populations. While generative models like GANs have been used to create synthetic data to fill these gaps, they often struggle with extreme ages due to data scarcity and suffer from "modal collapse," where they produce generic aging patterns that ignore individual diversity.

Proposed Method

The authors propose a style-conditioned generative framework that treats aging as a style transfer problem rather than a simple classification task. The architecture moves away from generic label-conditioning to capture fine-grained, person-specific aging traits.

Proposed style-based aging architecture featuring the Generator, Discriminator, and AdaIN style injection.

Architecture Components

  • The Generator: An autoencoder-based architecture that takes a source image and a target "style" reference image.
  • Style Injection via AdaIN: Using Adaptive Instance Normalization (AdaIN), the model extracts age-discriminative features from the target image at multiple resolutions. This allows the model to "paint" specific aging traits—like forehead wrinkles or gray hair—onto the source face.
  • Multi-task Discriminator: The discriminator learns to identify the age group while checking for realism, forcing the generator to produce statistically representative features.
  • Loss Functions: The framework utilizes a combination of Adversarial Loss (realism), Feature-matching Loss (style alignment), Reconstruction Loss (reversibility), and Identity-preservation Loss (maintaining the subject's identity).

Experimental Evaluation

The model was evaluated on three major benchmarks: MORPH (controlled environment), CACD (celebrity images "in-the-wild"), and FG-NET (used to test generalization to unseen data).

Key Results

The proposed method significantly outperformed baselines (CAAE and IPCGAN) in reaching target ages, particularly for the "50+" age group where previous models often failed to synthesize sufficient aging effects. Qualitatively, the model successfully transferred complex traits like receding hair lines and skin texture changes while maintaining the original person's identity.

Qualitative comparison of the proposed method against CAAE and IPCGAG across CACD, MORPH, and FG-NET datasets.

Beyond visual realism, the authors used ecological diversity metrics (Shannon and Simpson indices) to quantify the impact of their data augmentation. The results showed that the proposed method achieved a nearly "even" distribution across age groups, effectively rebalancing biased datasets more successfully than any compared method.

Conclusion

By treating age as a style, this framework captures the diverse ways individuals age, particularly in the "tails" of the age distribution. While current work focuses on age, future research aims to address attribute entanglement by handling multiple demographics (gender, race) simultaneously to further enhance algorithmic fairness. Scaling the current 128x128 resolution to high-definition imagery remains a standard challenge for future iterations.

Summary updated

Abstract

A significant limiting factor in training fair classifiers relates to the presence of dataset bias. In particular, face datasets are typically biased in terms of attributes such as gender, age, and race. If not mitigated, bias leads to algorithms that exhibit unfair behaviour towards such groups. In this work, we address the problem of increasing the diversity of face datasets with respect to age. Concretely, we propose a novel, generative style-based architecture for data augmentation that captures fine-grained aging patterns by conditioning on multi-resolution age-discriminative representations. By evaluating on several age-annotated datasets in both single- and cross-database experiments, we show that the proposed method outperforms state-of-the-art algorithms for age transfer, especially in the case of age groups that lie in the tails of the label distribution. We further show significantly increased diversity in the augmented datasets, outperforming all compared methods according to established metrics.
BibTeX
@inproceedings{georgopoulos2020enhancing,
  author = {Markos Georgopoulos and James Oldfield and Mihalis A. Nicolaou and Yannis Panagakis and Maja Pantic},
  title = {{Enhancing Facial Data Diversity with Style-based Face Aging}},
  booktitle = {Workshops @ CVPR},
  year = {2020},
  url = {https://openreview.net/forum?id=fQlzYCygHW6},
}
Figure