Person re-identification based on attention of fine-grained features

Abstract

In people re-identification tasks, the most intuitive method for selecting people representation features is to directly extract a global feature map of the people. However, relying on global features alone often fails to accurately identify people in the presence of occlusion, misalignment and background interference. In addition, due to interference from factors such as low camera resolution and illumination, some key local features (e.g. carried objects, body parts such as the face or limbs) are not clearly observed. To this end, we are inspired by fine-grained image classification tasks and propose an attentionbased fine-grained feature network model (AFGF) for people re-identification tasks, which will fully consider the relationship between global and local features and incorporate attention mechanisms to effectively extract fine features of people and improve their discrimination ability. The effectiveness of our model is validated on the Market-1501 and DukeMTMC-reID datasets, and the experimental results show that the method improves significantly on both the supervised baseline and unsupervised baselines.