A Global-Context Vision Transformer for Facial Age Estimation

Abstract Estimating a person’s age from a single facial image remains challenging because ageing cues are subtle, non-linear, and easily confounded by identity, pose, illumination, and demographic factors. In this work, we investigate the suitability of the Global Context Vision Transformer (GC-ViT), originally introduced for general vision tasks, as a pure transformer backbone for facial age estimation. We extend GC-ViT to scalar age regression through a lightweight MLP head and a task-specific training strategy, with the emphasis placed on a targeted reformulation of the architecture for facial age estimation and on its comprehensive empirical validation in this setting. GC-ViT is well suited to age estimation because it combines short-range window self-attention with long-range global attention, enabling a single hierarchical transformer to capture both fine texture cues (e.g., wrinkles) and global facial proportions. To improve robustness under limited data, we introduce an age-aware region replacement augmentation that replaces semantically defined facial regions with age-proximal donor content, or with self-flipped target content, and blends them using a feathered mask together with area-proportional soft label interpolation in donor mode. Training uses an ablation-validated composite regression loss (MAE with a secondary stabilising term, MSE or SmoothL1 depending on the dataset). We evaluate on MORPH II, IMDB-Clean, AFAD, and UTKFace, reporting MAE together with Pearson correlation and cumulative scores (CS@5/CS@10). Our GC-ViT achieves an MAE of 2.41 on MORPH II, 3.11 on AFAD, and 4.26 on UTKFace (the latter two obtained by fine-tuning from IMDB-Clean pretraining), while also being assessed under zero-shot cross-dataset transfer to quantify domain shift and out-of-domain generalization. Overall, the results show that an adapted local–global attention transformer, trained with targeted augmentation and a stable regression objective, can serve as a strong transformer-based baseline for facial age estimation without auxiliary CNN branches or test-time post-processing.