Unsupervised person re-identification by part-compensated soft multi-label learning

Abstract

Most existing re-ID models are developed based on supervised learning, which relies on plenty of manually labelled pairwise data for training. However, in many practical applications, such deep re-ID models are not scalable because of lacking sufficient annotated data. To address this problem, the superiority of unsupervised person re-ID methods have been reported and become more popular recently. In this paper, an unsupervised framework based on part-compensated soft multi-label learning is presented, which aims to explore the potential label information for unlabelled persons through mining both global and part level visual features. This re-ID framework is mainly composed of global clustering, part clustering and multi-label assignment. During clustering process, the results of part-level clustering are compensated into global clustering, allowing the network to learn more detailed information of pedestrians. Meanwhile, fuzzy clustering is used to generate soft multi-labels instead of commonly used single-label clustering. To further optimize the model, a cross-domain excitation parameter is introduced to relieve the background interference arising in the inter-domain view intersection. The proposed approach is evaluated on three public datasets, including Market-1501, DukeMTMC-reID, and MSMT17. Extensive experiments are carried out and this network can outperform the state-of-the-art unsupervised re-ID models.