Unsupervised Person Re-Identification with Attention-Guided Fine-Grained Features and Symmetric Contrast Learning

Unsupervised person re-identification has attracted a lot of attention due to its strong potential to adapt to new environments without manual annotation, but learning to recognise features in disjoint camera views without annotation is still challenging. Existing studies tend to ignore the optimisation of feature extractors in the feature-extraction stage of this task, while the use of traditional losses in the unsupervised learning stage severely affects the performance of the model. Additionally the use of a contrast learning framework in the latest methods uses only a single cluster centre or all instance features, without considering the correctness and diversity of the samples in the class, which affects the training of the model. Therefore, in this paper, we design an unsupervised person-re-identification framework called attention-guided fine-grained feature network and symmetric contrast learning (AFFSCL) to improve the two stages in the unsupervised person-re-identification task. AFFSCL focuses on learning recognition features through two key modules, namely the Attention-guided Fine-grained Feature network (AFF) and the Symmetric Contrast Learning module (SCL). Specifically, the attention-guided fine-grained feature network enhances the network's ability to discriminate pedestrians by performing further attention operations on fine-grained features to obtain detailed features of pedestrians. The symmetric contrast learning module replaces the traditional loss function to exploit the information potential given by the multiple samples and maintains the stability and generalisation capability of the model. The performance of the USL and UDA methods is tested on the Market-1501 and DukeMTMC-reID datasets by means of the results, which demonstrate that the method outperforms some existing methods, indicating the superiority of the framework.