Graph-Based Local Feature Adaptation for Cross-Domain Person Re-Identification

Abstract

The performance of cross-domain person re-identification has been greatly improved in recent years. However, there are still two problems in existing cross-domain person re-identification methods. First, most of them conduct domain adaptation on features that contain background noise. Second, they ignore the correlation between different features, including intra-domain and inter-domain. To overcome these problems, we propose a novel Graph-based Local Feature Adaptation (GLFA) framework for cross-domain person re-identification, which promotes domain adaptation by correlating intra-domain and inter-domain semantic local features with graph convolutional network. Specifically, in the feature extraction stage, we utilize a parsing model to extract semantic local features for both source and target domain samples, so as to avoid the feature shift caused by background discrepancy. In the feature adaptation stage, we apply two stacked Graph Convolutional Networks (GCN) to propagate semantic local feature information within each domain and across different domains respectively. It promotes the transfer of specific knowledge from the source domain to the target domain. Furthermore, to ensure the features from different domains updated by GCN are well aligned, a local feature distribution alignment loss is introduced on the top of GCN. The combination of GCN and alignment loss enables our framework diminish the feature shift caused by other factors effectively. Extensive experiments on Market-1501, DukeMTMC-reID and MSMT17 datasets demonstrate that our method outperforms the state-of-the-art methods evidently.