Gu Xinqian ; Ma Bingpeng ; Chang Hong ; Shan Shiguang ; Chen Xilin

Abstract

In many scenarios of Person Re-identification (Re-ID), the gallery setconsists of lots of surveillance videos and the query is just an image, thusRe-ID has to be conducted between image and videos. Compared with videos, stillperson images lack temporal information. Besides, the information asymmetrybetween image and video features increases the difficulty in matching imagesand videos. To solve this problem, we propose a novel Temporal KnowledgePropagation (TKP) method which propagates the temporal knowledge learned by thevideo representation network to the image representation network. Specifically,given the input videos, we enforce the image representation network to fit theoutputs of video representation network in a shared feature space. With backpropagation, temporal knowledge can be transferred to enhance the imagefeatures and the information asymmetry problem can be alleviated. Withadditional classification and integrated triplet losses, our model can learnexpressive and discriminative image and video features for image-to-videore-identification. Extensive experiments demonstrate the effectiveness of ourmethod and the overall results on two widely used datasets surpass thestate-of-the-art methods by a large margin. Code is available at:https://github.com/guxinqian/TKP