Abstract :
Human age, gender and ethnicity are valuable demographic information about a population. These measures are also considered important soft biometric traits for human recognition or search. Usually the three traits are studied separately. A recent study [9] shows that the three traits can be estimated simultaneously based on a multi-label regression formulation. The linear and kernel partial least squares (PLS) models are adopted to solve the multi-label regression problem in [9]. In this study, we investigate the canonical correlation analysis (CCA) based methods, including linear CCA, regularized CCA (rCCA), and kernel CCA (KCCA), and compare to the PLS models in solving the joint estimation problem. Interestingly, we found a consistent ranking of the five methods in estimating age, gender, and ethnicity. More importantly, we found that the CCA based methods can derive an extremely low dimensionality in estimating age, gender and ethnicity, which has not been shown in previous research, to the best of our knowledge. The experiments are conducted on a very large database of more than 55,000 face images.
Keywords :
age issues; biometrics (access control); correlation methods; face recognition; gender issues; least squares approximations; regression analysis; KCCA; PLS models; age joint estimation; canonical correlation analysis based method; demographic information; ethnicity joint estimation; face images; gender joint estimation; human recognition; human search; kernel CCA; kernel partial least squares model; linear CCA; linear kernel partial least squares model; multilabel regression formulation; rCCA; regularized CCA; soft biometric traits; Accuracy; Databases; Eigenvalues and eigenfunctions; Estimation; Feature extraction; Kernel; Vectors;
