Enhanced Deep Feature Representation for Person Search

Abstract

Person re-identification (re-id) has attracted widespread attention due to its application and research significance. However, since the person re-id puts the cropped images as input, it is far from the real-world scenarios just like person search which aims at matching a target person from a gallery of the whole scene images. Person search is more difficult but more practical and meaningful. In this paper, we propose a new person search network with an enhanced feature representation. Our network mainly consists two parts, a pedestrian proposal net and an identification net. In the identification net, we utilize hand-crafted features and Convolutional Neural Network (CNN) features to get more discriminative and compact features. Experiments on a large-scale benchmark dataset demonstrate our network gets better performance than others counterparts.