Unsupervised Multi-Source Domain Adaptation Person Re-Identification Methods Based on Transformer and Graph Convolutional Neural Network

The Person Re-Identification expects to retrieve the unique pedestrians across multiple non-overlapping cameras. Mainstream unsupervised person Re-Identification methods generate pseudo-labels by clustering in single-source and single-target domain scenarios. Although convolution neural network (CNN)-based methods have achieved promising progress, existing methods loss some feature information during convolution and downsampling operators. Moreover, due to the dramatic disparities between different source datasets, the model trained on the multi-source domain directly will lead to poor performance. To overcome these limitations, we propose a unsupervised multi-source domain adaptation person Re-Identification method based on transformer and graph convolutional neural network. Firstly, we introduce a pure transformer framework for Re-ID task to improve the model's feature representation capability. Secondly, we suggest using domain information fusion module based on graph convolutional neural network to minimize the gap among the source domains, and employ a feature enhancer based on Gaussian distribution to reduce domain-specific characteristics and increase the distinctiveness of person features. Finally, extensive experiments on four benchmark datasets show that the proposed method outperforms most of the current unsupervised domain adaptation person Re-Identification methods.