DSFNet: Dual-fusion network with secondary clustering and feature integration for unsupervised person re-identification

Unsupervised person re-identification (ReID) aims to train a model by updating a memory dictionary to retrieve a person of interest from different camera views. This area has attracted widespread interest recently due to its low cost of data processing. Existing methods predominantly concentrate on utilizing a single, fixed optimization approach (network) for clustering all features, neglecting their diversity and integrity, thereby resulting in noisy pseudo-labels. To solve this problem, this study proposes a DSFNet framework, namely, a Dual-fusion network with secondary clustering and feature integration (DSFNet) framework for unsupervised person ReID tasks. The proposed framework consists of three main components: (i) a hard-sample secondary clustering network (SCNet), (ii) a feature integration network (FINet), and (iii) a Dual-fusion dynamic optimization (DDO) scheme. Specifically, the first module focuses on initializing the memory dictionary via a hard-sample (dissimilar samples in the intra-class or similar samples in the inter-class) secondary clustering strategy, which preserves the diversity of the individual features. The FINet explores each person’s local and global features via an integrated weight-assigned strategy to ensure the integrity of the individual instance features. Next, the dual-fusion dynamic optimization scheme is implemented from FINet to SCNet, thereby guaranteeing consistent clustering features and ultimately mitigates the noise related to the generation of pseudo-labels. Extensive experimental results demonstrate that, compared to the latest pure unsupervised methods on commonly used ReID datasets (such as Market-1501, MSMT17, and PersonX), our approach achieves performance improvements in mAP of 1.7%, 1.9% and 4.0%, respectively. Similarly, Rank@1 performance has seen improvements of 0.1%, 4.6% and 1.0%, respectively. Meanwhile, to verify the generalization capability of our method, we conducted tests on the additional vehicle re-identification dataset VeRi-776, which exhibited performance largely comparable to the latest methods.