Uncertainty-guided Robust labels refinement for unsupervised person re-identification.

Existing state-of-the-art unsupervised person re-identification (Re-ID) methods rely on clustering to generate pseudo labels for training. The reliability of pseudo labels directly affects the performance of these methods. Due to unsatisfactory feature embedding and imperfect clustering, pseudo labels are not always reliable that significantly hinder the representation learning of person. To address this issue, an Uncertainty-guided Robust labels refinement method is proposed. It employs uncertainty as guidance to selects and optimizes unreliable pseudo labels, which alleviates the impact of incorrect pseudo labels. Specially, an uncertainty estimation module is constructed, which identifies the samples with low reliability applies base on consistency between the ideal distribution and the predicted distribution. To mitigate the impact of unreliable samples, the labels refinement module is designed, which rebuilds labels for low reliability samples by measuring similarity from the closest centroid. Thanks to the reliability of pseudo labels provided by uncertain estimation module, the proposed method enhances robustness to noisy labels and learns discriminative representations of person. Extensive experiments demonstrate that the proposed method is effective and achieves advanced performance for unsupervised person Re-ID. [ABSTRACT FROM AUTHOR]