A Novel Attention-Driven Framework for Unsupervised Pedestrian Re-identification with Clustering Optimization

Abstract

Unsupervised pedestrian re-identification (re-ID) is not merely a visual recognition task; it represents a significant sub-field within the domain of pattern recognition. Despite the remarkable success of Convolutional Neural Networks (CNN) in re-ID, they still face challenges in handling variations in pose, occlusion, and lighting conditions. To effectively tackle these challenges, it is imperative to prioritize implementing efficient sampling strategies. We propose a Novel Attention-Driven Framework for Unsupervised Pedestrian re-ID with Clustering Optimization (AFC) to address the above issues. First, we introduce a new attention mechanism that enhances multi-scale spatial attention and reduces the number of trainable parameters. Then, we employed a straightforward and effective method of group sampling. In addition, we apply a clustering consensus approach to estimate pseudo-label similarity in continuous training and use temporal propagation and ensembles to improve pseudo-labels. Extensive experiments on Market-1501, duketmc-reID and MSMT17 datasets show that our method achieves significant performance improvement in unsupervised pedestrian re-ID, which provides important theoretical and practical value for the research on deep fusion of pattern recognition field with pedestrian re-ID and promotes the further development of the related fields.