Adaptive Graph Attention Network in Person Re-Identification

Abstract

This paper approaches a significant problem in computer vision: re-identifying a person when having groups of people. Re-identifying by group context is a new direction for improving traditional single-object re-identifying tasks by additional information from group layout and group member variations. Furthermore, adding new improvements in the graph convolution layer structure or using more powerful theories enhances the model's accuracy. In this study, we propose to leverage the information of group objects: people and subgroups of two or three people inside a group image from the CUHK-SYSU dataset. The organization of data is based on the relational representation of the central node, and the observed nodes further incorporate their features extracted through the Resnet backbone. We also recommend using the SeLU activation function in the graph convolution model for experiments. The key challenge in implementing is to define the optimal group-wise matching using adaptive graph attention based on a graph convolution network modified and training techniques. The experiment results showed that our method improved the model's learning efficiency by approximately 1.2% compared to the mean average precision score. Moreover, the optimal number of learning parameters is reduced to one third compared to the original.