Quality Guided Metric Learning for Domain Adaptation Person Re-Identification

Abstract

Person re-identification is the task of identifying pedestrians across different cameras. Domain adaptation person re-identification involves transferring knowledge from labeled source domains to unlabeled target domains, with applications in security and surveillance. Challenges emerge due to variations in sample quality and disparities in distance distribution between positive and negative sample pairs. To address these challenges, this paper proposes a quality guided metric learning approach for domain adaptation person re-identification. We focus on improving appearance similarity metrics by evaluating sample quality based on local visibility, categorizing images as high or low quality. Besides, we introduce an adaptive weight triplet loss incorporating camera information to optimize triplets. This reduces the effects of invalid triplets and facilitating ongoing target domain learning.We have conducted comprehensive comparative evaluations to showcase the advantages and superiority of our proposed method. Our method has 2.6%, 1.9%, and 6.2% improved on Market-1501, DukeMTMC-reID, and MSMT17 datasets, respectively.