Occluded Person Re-Identification by Multi-Granularity Generation Adversarial Network

Abstract

In order to address the problem that the detailed features of pedestrians are not prominent and the pedestrian pictures are obscured in unique environments in the process of person re-recognition, we propose a person re-recognition method with a multi-grain size generative adversarial network. Firstly, we use the generative adversarial network to recover the occluded pedestrian pictures; secondly, we improve the traditional multi-granularity network by adding an Efficient Channel Attention for Deep Convolutional Neural Networks (ECA-Net) on the coarse-grained branch to focus on the feature information in the pedestrian pictures and use the High-Resolution Net (HRNet) for pose estimation on the fine-grained branch to divide the pedestrian pictures into nine parts, to enhance the network’s learning of more detailed features of pedestrians, and thus improve the accuracy of pedestrian re-recognition learning, which in turn improves the accuracy of person re-identification.

Keywords

• Person re-identification
• generative adversarial networks
• random occlusion
• attention mechanism