Make the Bag Disappear: Carrying Status-invariant Gait-based Human Age Estimation using Parallel Generative Adversarial Networks

Abstract

Existing approaches to gait-based human age estimation seldom consider variations such as carried objects, which greatly alter appearance of gait features (e.g., gait energy images) and result in poor age estimation results. Therefore, we propose a method of gait-based human age estimation robust against carrying status using generative adversarial networks. Specifically, we consider a generative network that outputs a gait feature without carried objects (i.e., it makes the carried objects disappear) given an input gait feature with or without carried objects, and then learns the parallel generative network with shared weights from a pair of gait features of the same subject with and without carried objects. Thereafter, the generated gait features are separately fed into the same subsequent age regression network for age estimation, which is trained in an end-to-end manner in conjunction with the parallel generative network. Because of this design, the proposed method can generate gait features without carried objects that are suitable for gait-based age estimation regardless of the carrying status of the input gait features, and therefore improve the age estimation accuracy under carrying status variations. Experimental results on a large gait dataset with carried objects show the state-of-the-art performance of the proposed method.