DCLR-SF: distribution consistent label refinement and lighten similarity network fusion for multi-source domain-adaptive person re-identification

Unsupervised domain-adaptive person re-identification is challenging and has received much attention in recent years. Some previous works generate pseudo-labels by clustering algorithms to further optimize the model. But pseudo-labels inevitably contain noise due to the domain shift and the imperfection of the clustering methods. This article attempts to solve this problem by proposing two solutions: cross-supervision and distribution consistent label refinement (DCLR) schemes. For the former, we use labeled multi-source domain data to enhance the complementarity of different branches in the network and then exchange pseudo-labels for suppressing the negative effects of noise. For the latter, the proposed DCLR enlarges the pairwise distance of different identity features in a coarse cluster and then uses density peaks clustering technology to refine. This directly improves the quality of pseudo-labels. In addition, we design an adaptation learning-based memory bank to improve the consistency of attributes across domains and reduce the domain shift. This study also proposes the lighten similarity network fusion for person re-identification tasks to fuse the query results of multiple branches in the network. Extensive experiments demonstrate the superiority of our method on multiple person re-identification datasets.