Graph Convolutional Network for Person Re-Identification Based on Part Representation

Local expression based on a pedestrian graph is an effective way to solve many problems in the person's re-identification. There is, however, a certain spatial relationship between the various local features of pedestrian graphs. This paper proposes a convolutional graph based on local pedestrian features of the part-based graph constructive network (PB-GCN) learning framework to address the above issues. The framework constructs a topological relationship between global and local features, obtains a new feature representation through a graph conversion network, and then trains and tests the representation of features. By comparison of single and multiple query results on the Market1501 dataset, PB-GCN (ResNet-50) achieves 94.1 and comparison of models on DukeMTMC-reID and MSMT17 datasets, PB-GCN + re-rank (ResNet-50) attains 79.6.