Leveraging Teacher Ensembles to Mitigate Label Noise in Clustering-Based Unsupervised Domain Adaptation for Person Re-Identification

Person re-identification (re-ID) pertains to the task of establishing correspondences between individuals across disparate camera views and holds relevance in the fields of surveillance and security. Recently, unsupervised domain adaptation (UDA) received lots of attention for person re-identification (re-ID). The prime objective of UDA is to enhance the performance of models by transforming knowledge of the source domain to the target domain, reducing the dependence on labeled data. This paper proposes an efficient approach that leverages teacher ensembles to mitigate label noise in clustering-based UDA for person re-ID. The proposed method involves transferring the knowledge from multiple teacher models to a student model, enabling effective adaptation while minimizing the reliance on labeled target domain data. Extensive experiments are conducted in prominent re-ID datasets to evaluate the impact of different factors, such as teacher network architecture and ensemble size variation for the proposed scheme. From the experimental results, it has been noticed that a well-selected teacher ensemble and an optimal ensemble size significantly contribute to improving the performance of clustering-based UDA. The proposed scheme brings more competitive performance gains than that of the state-of-the-art re-ID approaches. Further, this work contributes to the advancement of UDA techniques in person re-ID by addressing the challenges posed by label noise and showcasing the potential of teacher ensembles in enhancing clustering-based UDA performance.