LIAAD: Lightweight attentive angular distillation for large-scale age-invariant face recognition

Disentangled representations have been commonly adopted to Age-invariant Face Recognition (AiFR) tasks. However, these methods reached some limitations with (1) the requirement of large-scale face recognition (FR) training data age labels, which is limited in practice; (2) heavy deep network architectures for high performance; and (3) their evaluations are usually taken place on age-related databases while neglecting standard FR guarantee robustness. This work presents a novel Lightweight Attentive Angular Distillation (LIAAD) approach Large-scale AiFR that overcomes limitations. Given two high-performance networks as teachers different specialized knowledge, LIAAD introduces learning paradigm efficiently distill age-invariant attentive angular knowledge from those lightweight student making it more powerful higher accuracy robust against factor. Consequently, able take advantages both datasets without labels train an model. Far apart prior distillation mainly focusing compression ratios closed-set problems, our aims solve open-set problem, i.e. recognition. Evaluations LFW, IJB-B IJB-C Janus, AgeDB MegaFace-FGNet one million distractors demonstrated efficiency proposed light-weight structure. also new longitudinal aging (LogiFace) database 1 further studies facial problems future.