Hybrid feature constraint with clustering for unsupervised person re-identification

Unsupervised person re-identification (Re-ID) has better scalability and usability in real-world deployments due to the lack of annotations, which is more challenging than supervised methods. State-of-the-art approaches mainly employ clustering algorithms to generate pseudo-labels for transferring the process into a supervised operation. However, the clustering algorithm depends on discriminative pedestrian features. Only using the clustering algorithm produces low-quality labels and hinders the performance of the Re-ID model. In the paper, we propose the hybrid feature constraint network (HFCN) to adequately restrict the pedestrian feature distribution for unsupervised person Re-ID. Specifically, we first define a feature constraint loss to restrict the feature distribution so that different pedestrians can be clearly distinguished at the first step. And then, we design a multi-task operation with the iterative update for clustering algorithm to further implement the feature constraint. This can adequately utilize predicted label information and identify complex samples. Finally, we integrate the feature constraint loss and multi-task operation to optimize the Re-ID model, which could promote the clustering to generate high-quality labels and learn valuable information. Extensive experiments prove that the proposed HFCN is effective and outperforms the state-of-the-art.