Part-Based Feature Complementary Denoising for Unsupervised Person Re-Identification

Unsupervised Person re-identification (ReID) aims to automatically capture and match images of the same person across different camera viewpoints without any manual annotations. Current methods primarily generate pseudo-labels by clustering global features and employ contrastive learning strategies for training. Despite the promising advancements made by these approaches, effectively addressing the inherent bias of global features and mitigating the impact of pseudo-label noise remains an unresolved issue. To tackle this challenge, we propose a part-based features complementary denoising method (PFCD). Specifically, we design the combined features (CF) module and the partial-features fusion and contrastive scheme (PFCS), which capture fine-grained clues from a local perspective and combine global and local features for clustering with consistent pseudo-label assignment, thereby achieving a complementarity between global and local features. Furthermore, to diminish the influence of pseudo-label noise on the model, we design the GMM features denoising (GFD) module, which employs a Gaussian Mixture Model to categorize features within each pseudo-class based on confidence levels and performs denoising on low-confidence features. Lastly, we construct a modular knowledge distillation (MKD) to enhance feature representation capabilities and effectively reduce pseudo-label noise. In addition, our method is confirmed to be effective through extensive experiments on four challenging ReID datasets, remarkably surpassing numerous state-of-the-art methods. Code has been made available at https://github.com/xfltdzzz/PFCS_ReID.