Multi-loss Function-based GAN for Cross-view Gait Recognition

Abstract

Gait is a very promising biometrics. However, a great challenge in this area is how to improve the recognition accuracy in cross-view settings. In this study, a multi-loss-based GAN (MLF-GAN) was proposed for gait transformation between arbitrary views and then for view-consistent identity recognition. The generation of gaits was regularized using an identity preserver together with a discriminator. The discriminator comprises of two components, whose network structures are the same except the last layer. One component is for judging whether the generated images are realistic, and the other one is used for ensuring the view consistency between the generated and the target gaits. To better retain identity information during gaits transformation, the identity preserver also utilize two stacked components, where one is optimized by triplet loss and the other one is optimized by cross-entropy loss. The distribution of the latent gait features from the global perspective is regularized by the cross-entropy loss, while the fine-grained local features that are beneficial for classification are learned by way of the triplet loss. Experimental results on CASIA-B gait database demonstrate the effectiveness of the proposed method, and the comparison with the state-of-the-art indicates that our method contributes to the accuracy improvement of cross-view gait recognition.