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

The performance of facial analysis algorithms is heavily dependent on the quality and diversity of the training data. However, most large-scale facial datasets exhibit significant demographic biases, particularly regarding age. For instance, the FG-NET dataset is skewed, with 87% of samples representing individuals under 30. Such biases lead to models that generalize poorly to underrepresented age groups, such as the elderly. While Generative Adversarial Networks (GANs) offer a potential solution through data augmentation, existing methods often suffer from "mode collapse," producing generic aging patterns that lack the fine-grained diversity seen in real-world aging. This work introduces a style-based approach to synthesize diverse, realistic facial data to mitigate these biases.

Proposed Method

We propose a Style-conditioned Generator that treats age as a "style" rather than a discrete category. Unlike standard GANs that condition on a simple label, our model uses a target image to define the specific aging patterns to be transferred.

Overview of the proposed style-based face aging architecture.

As illustrated in Figure 1, the architecture utilizes multi-resolution style injection. A discriminator extracts age-discriminative features (mean and standard deviation of activations) from a target image. These features are then injected into the generator's decoder at multiple scales using Adaptive Instance Normalization (AdaIN). This allows the model to capture both global structural changes (e.g., face shape) and fine-grained textures (e.g., specific wrinkle patterns). The training is guided by a combination of adversarial loss, feature-matching loss ($L_{fm}$), cycle-consistency loss, and identity preservation loss ($L_{id}$).

Experimental Evaluation

The model was evaluated on three major benchmarks: MORPH, CACD, and FG-NET. Our approach demonstrates superior aging accuracy and diversity across all datasets.

Qualitative results showing face aging across different age categories.

Quantitative Results

On the MORPH dataset, the proposed method achieved a mean absolute age difference of only 1.69 years for the 50+ age group, significantly outperforming baselines such as CAAE (12.96 years) and IPCGAN (4.03 years). Furthermore, diversity analysis using Shannon and Simpson indices confirms that our method produces more balanced and varied augmented datasets compared to state-of-the-art alternatives.

Qualitative Analysis

Visual results (Figure 2) show that while baselines often produce blurry or over-regularized faces, our method preserves sharp identity details while adding realistic, target-specific aging textures. By leveraging the discriminator's learned feature representations, the model can generate a wide variety of aging outcomes for the same individual, reflecting real-world intra-class diversity.

Conclusion

This paper presents a novel style-based GAN architecture for face aging that effectively enhances dataset diversity. By moving beyond label-based conditioning and treating age as a style, we mitigate mode collapse and produce high-fidelity synthetic data. While the current framework focuses on age, future work will explore generalizing this approach to other demographic attributes like gender and race to further improve algorithmic fairness in facial analysis.

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