Camera-aware cluster-instance joint online learning for unsupervised person re-identification

Unsupervised person re-identification (re-ID) aims at learning discriminative feature representations for person retrieval without any annotations. Pseudo-label-based methods that iteratively perform pseudo-label generation and model training are currently the most popular approach to achieve this goal. However, distribution variations among cameras inevitably introduce noise in the generated pseudo-labels. Moreover, they are often assigned offline using relatively simple clustering criteria, which further accumulates the noise and limits the potential improvement in model performance. To address these issues, we propose a novel camera-aware cluster-instance joint online learning (CCIOL) framework that leverages the online inter-camera K-reciprocal nearest neighbors (OICKRNs) mined for each sample at every iteration to soften the traditional hard pseudolabels at the cluster-level and generate multi-labels at the instance-level. Additionally, contrastive learning losses at two levels are employed to rectify the erroneous closeness between samples and promote intraclass aggregation and inter-class separation. Extensive experimental results on Market1501 and MSMT17 demonstrate the competitiveness of the proposed method compared to state -of -the -art unsupervised re-ID approaches.