Single view facial age estimation using deep learning with cascaded random forests.

The task of estimating a person’s real age using unconstrained facial images has been actively studied in biometrics research. We developed several deep learning architectures and supervision methods for facial age estimation and evaluate the impact of different pre-processing and face alignment (or normalization) methods on the feature embedding subspace. The proposed novel two-stage supervised learning model utilizes ResNeXt as a backbone combined with a two-layer random forest (TLRF) to estimate age. Our deep architectures are trained using a custom loss function to handle variations in gender, pose, illumination, ethnicity, expression and context, on the VGG-Face2 MIVIA Age Dataset with over 575K images, as part of the Guess the Age (GTA) contest. Surprisingly, face alignment using FANet during training did not improve accuracy. We were able to achieve an Age Accuracy and Regularity score AAR = 7.02 with a variance σ = 1.16 using only ResNeXt. The proposed ResNeXt+TLRF model improved age-class generalizability with a smaller variance of σ = 0.98 and a second best AAR = 6.97.