Task-Specific Neural Networks for Pose Estimation in Person Re-identification Task

Abstract

Person re-identification is a challenging task because of severe appearance changes of a person due to diverse camera viewpoints and person poses. To alleviate the impact of different poses, more and more studies have been done in pose estimation. In this work, we present three task-specific neural networks (TNN) algorithm designed to address the problem of pose estimation for re-identification in both single-shot and multi-shot matching. In order to recognize the human pose as one of the four classes (front, back, left, right), a PoseNet-A is first required to estimate the pose as class front-back or class left-right. Based on the results, we select the appropriate network (PoseNet-B1, PoseNet-B2) to obtain the final pose. According to the results, our method achieves very good results on a large data set (CUHK03-Pose). One thing that needs to be pointed out is that we build the dataset CUHK03-Pose which is based on the person re-identification dataset CUHK03.