Face Recognition Using Shallow Age-Invariant Data

Face aging is a challenging research area in face recognition and verification. However, recognizing identical faces with an immense age gap gives rise to various difficulties because of the diversity in face appearance caused by aging. In real-world applications, the pre-training dataset is very limited (e.g only two age-invariant face images) are present for each identity. We define this term as Shallow Age-Invariant Face Learning, and find it difficult with existing methods. Different from large-scale and shallow face data, the age-invariant dataset is lacking intra-class distance. Thus, mostly face recognition algorithms are collapsed, feature dimensions easily gets into over-fitting and degeneration problems. To examine such discrepancies and evaluate recent state-of-the-art methods (SOTA), we investigate the performance of SOTA models on different metrics learning approaches. Further, we re-organize three age-invariant datasets into negative and positive pairs with a mixed interval of ages. We select the face aging dataset (FG-NET) to cover all age groups with aging and occlusion, the longitudinal dataset (MORPH-Album 2) for illumination and low-resolution, and the age database (Age-DB) for pose and noise. Although for major factors, Arc-Face and Face-Net outperform others. Another interesting finding of our research, the training performance on age-invariant pairs is still saturated with a clear margin. This research provides further directions for future research and insights.