Zhang Yunxuan Liu Li Li Cheng Loy Chen change

Abstract

We introduce a novel approach for annotating large quantity of in-the-wildfacial images with high-quality posterior age distribution as labels. Eachposterior provides a probability distribution of estimated ages for a face. Ourapproach is motivated by observations that it is easier to distinguish who isthe older of two people than to determine the person's actual age. Given areference database with samples of known ages and a dataset to label, we cantransfer reliable annotations from the former to the latter viahuman-in-the-loop comparisons. We show an effective way to transform suchcomparisons to posterior via fully-connected and SoftMax layers, so as topermit end-to-end training in a deep network. Thanks to the efficient andeffective annotation approach, we collect a new large-scale facial age dataset,dubbed `MegaAge', which consists of 41,941 images. Data can be downloaded fromour project page mmlab.ie.cuhk.edu.hk/projects/MegaAge andgithub.com/zyx2012/Age_estimation_BMVC2017. With the dataset, we train anetwork that jointly performs ordinal hyperplane classification and posteriordistribution learning. Our approach achieves state-of-the-art results onpopular benchmarks such as MORPH2, Adience, and the newly proposed MegaAge.