| Abstrakt: |
Unsupervised person re-identification (re-ID) aims to identify the same person in unsupervised settings, which is a realistic and challenging problem due to getting rid of the dependence on annotations. Existing pseudo-label-based methods have been proved to be effective by discriminative representation learning and high-quality pseudo labels. However, there are some undesirable similarity structures in images affected by different states (e.g., backgrounds, occlusions, or other target confusions) which may generate noisy labels and lead to model degradation. To tackle the above issues, this paper analyzes the key factors for unsupervised re-ID performance from two perspectives: network structure and loss function. Specifically, we propose a multi-confidence contrastive learning with feature refinement (MCFR) model which can enhance the feature discriminative ability and meanwhile alleviate the impact of noisy data. For network architecture, a feature refined block is embedded into a multi-branch network, constructing FR-net. For loss function, a multi-confidence contrastive loss (MCCL) based on similarity and confidence relationship is developed. MCCL can flexibly update the list of hard negative samples in clusters and outliers, promoting the model by learning from confident samples. Extensive experiments including visualizations and ablation studies validate the effectiveness of each component in the MCFR model. Additionally, compared with the state-of-the-art unsupervised re-ID methods, our method achieves considerable performance on three benchmark datasets. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink