Transferring discriminative knowledge via connective momentum clustering on person re-identification. (June 2022)

• Record Type: Journal Article
• Authors: Lu, Yichen; Deng, Weihong
• Is Part Of: Pattern recognition. Volume 126(2022)
• Publication Date: 2022-06
• Subjects: Person re-identification -- Unsupervised domain adaptation -- Graph convolutional networks -- Momentum mechanism -- Batch normalization; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108569
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Unsupervised domain adaptation (UDA) in person re-identification (ReID). Connective Momentum Clustering (CMC) framework to build a connection estimator via graph convolutional networks (GCN). Normalize the data stream separately to decouple different distribution. Extensive cross-database experiments on Duke and Market databases. SOTA performance on ReID domain adaption. Abstract: Unsupervised domain adaptation in person re-identification remains a challenge to learning discriminative representations due to the absence of labels in target domain. Clustering could provide pseudo-labels, but the limitation mainly comes from imperfect clustering and noisy pseudo-labels. To address this drawback, we propose C onnective M omentum C lustering (CMC) framework to build a connection estimator via graph convolutional networks to transfer rich connection knowledge from the annotation space of source data to target domain. It estimates connections from context to reveal relationship between unlabeled data and helps to discover more reliable clusters. With momentum mechanism, stable pseudo-labels are updated iteratively with confidence and refined consistently to encourage more discriminative networks. Meanwhile, we notice that the huge domain gap between source and target domains results in severe pollution in BatchNorm layers. To tackle this problem, we normalize the data stream separately to decouple different distribution and further boost the performance in target domain. We adopt our CMC framework on mainstream tasks and achieves 80.2% mAP / 91.3% Rank-1 on Duke → Market task and 70.4% mAP / 82.4% Rank-1 on Market → Duke task.