PrGCN: Probability prediction with graph convolutional network for person re-identification

• Published in: Neurocomputing (Amsterdam) 2021-01, Vol.423, p.57-70
• Main Authors: Liu, Hongmin, Xiao, Zhenzhen, Fan, Bin, Zeng, Hui, Zhang, Yifan, Jiang, Guoquan
• Format: Article
• Language: English
• Subjects: Graph convolutional network; Link probability prediction; Person re-identification; Similarity measurement
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2020.10.019

•Similarity computation in ReID is casted to a probability prediction problem with GCN.•The proposed PrGCN can be embedded into existing works to achieve better performance.•By combining PrGCN with PCB, state-of-the-art results are obtained on ReID benchmarks. Robust similarity measurement is an important issue for person re-identification (ReID). Most existing ReID models estimate the similarity between query and gallery images by computing their Euclidean distances while ignoring the rich context information contained in the image space. In this paper, we propose a graph convolutional network (GCN) based method to improve the similarity measurement in ReID, which regards the ReID task as a prediction problem of the link probability between node pairs. Our method is named as PrGCN (Probability GCN), in which each person is regarded as an instance node. Firstly, an Instance Centered Sub-graphs (ICS) is constructed for each instance node to depict its rich local context information. Secondly, the constructed ICS is input to a GCN to infer and predict the link probability of node pairs, followed by a similarity ranking between the query and gallery images according to the predicted probabilities. Extensive experiments show that the proposed method improves the mAP and Top-1 accuracy of ReID significantly, yielding better or comparable results to the state-of-the-art methods on various benchmarks (Market1501, DukeMTMC-ReID and CUHK03). In addition, we validate that the proposed PrGCN can be easily embedded into other deep learning architectures to replace Euclidean distance metric and achieve significant performance improvements.