Self-attention mechanism in person re-identification models

Abstract

In recent years, person re-identification based on video has become a hot topic in the field of person re-identification. The self-attention mechanism can improve the ability of deep neural networks in computer vision tasks such as image classification, image segmentation and natural language processing tasks. In order to verify whether the self-attention can improve the performance or not in person re-identification tasks, this paper applies two self-attention mechanisms, non-local attention and recurrent criss-cross attention to person re-identification model, and experiments are conducted on Market-1501, DukeMTMC-reID and MSMT17 person re-identification datasets. The results show that the self-attention mechanism can improve the accuracy of the person re-identification model. The accuracy is higher when the self-attention module is inserted into the convolutional layers of the re-identification network.