Within-network ensemble for face attributes classification

Abstract

Face attributes classification is drawing attention as a research topic with applications in multiple domains, such as video surveillance and social media analysis. In this work, we propose to train attributes in groups based on their localization (head, eyes, nose, cheek, mouth, shoulder, and general areas) in an end-to-end framework considering the correlations between the different attributes. Furthermore, a novel ensemble learning technique is introduced within the network itself that reduces the time of training compared to ensemble of several models. Our approach outperforms the state-of-the-art of the attributes with an average improvement of almost 0.60% and 0.48% points, on the public CELEBA and LFWA datasets, respectively.