Ensemble knowledge distillation for collaborative pseudo-label refinement in unsupervised domain adaptation for person re-identification

Abstract

In open-world person re-identification, obtaining annotated ground truth for every deployment environment is infeasible which necessitates the exploration of unsupervised domain adaptation methods. The Pseudo-label guided approaches leverage clustering methodologies for the generation of pseudo-labels to facilitate model retraining on the unlabeled target domain. However, the robustness of these methodological approaches is vulnerable in the presence of label noise, which is generated due to the domain shift, specifically, the interplay of intra-domain similarity within the source domain and the inter-domain dissimilarity between the source and target domains. This study introduces a novel learning scheme to enhance the robustness of such cross-domain re-identification models by minimizing the label noise through collaborative refinement utilizing the amalgamation of multiple teacher outputs and subsequent network training with the refined pseudo-labels. To obtain diverse perspectives, domain-adapted teacher networks of varying types and sizes are employed, and subsequently, their outcomes are amalgamated through a dynamic averaging method to yield a refined representation. A comprehensive set of ablation studies systematically analyzes the contributions of each component, highlighting the vital roles of clustering, knowledge distillation, collaborative refinement, and optimal ensemble size. Furthermore, the proposed model undergoes rigorous evaluation on widely used datasets, such as Market1501, DukeMTMC, and MSMT17, demonstrating superior performance compared to existing state-of-the-art methods. The mean Average Precision (mAP) and Rank-1 (R1) scores showcase notable improvements, with results indicating an enhanced performance in Duke to Market (mAP: 81.04, R1: 90.84) and Market to Duke (mAP: 76.24, R1: 86.27) scenarios.