SanNet: A Neural Network for Multitasking Face Attributes

抄録

We present a divide-and-conquer approach to simultaneously predicting the age, gender and mask status of a person from their head-face image. To this end, we propose a novel neural network architecture which for convenience, is called SanNet, hereinafter. At a high level, SanNet consists of a shared replaceable backbone, followed by three separate branches, namely; A-branch, G-branch and M-branch, for the age task, the gender task and the mask task, respectively. This architecture is inspired by the multitasking capacity of the human brain. The A-branch in SanNet performs regression and returns a prediction for the age group of the individual, while the G- and M-branches are binary classifiers. We perform experiments with different backbone architectures and using a public dataset, augmented for our purpose. Our preliminary results show that lighter models can achieve high accuracy for G- and M- branches, while heavier model is provides better MAE for A- branch.