A partial least squares based ranker for fast and accurate age estimation

Facial age estimation is challenging due to complex dynamics in aging process, which render metric regression methods unfavorable. Rankers show better performance by exploiting the ordinal nature of ages. The difficulty of designing a ranker is that each binary classifier of a ranker has to be trained using highly unbalanced positive and negative data. This paper proposes a partial least squares based ranker (PLS-Ranker), which fully maintains the advantages of PLS and greatly boosts its performance on the ordinal problem. In PLS-Ranker, an adaptive threshold learning strategy is proposed to boost each of the binary classifiers learned from highly unbalanced data. Previous ranking approaches such as CS-OHRank suffer from heavy computations because dozens of binary classifiers are trained separately. However, in PLS-Ranker, they are jointly learned. Additionally, PLS-Ranker simultaneously reduces feature dimensions and ranks in high speed even for high-dimensional features. Experimental results on the age estimation problem show that PLS-Ranker outperforms the state-of-the-art methods in terms of both accuracy and speed. PLS-Ranker also achieves state-of-the-art performance on the multi-source cross-race-and-gender age estimation problem, which further demonstrates its robustness.