1. A plug-and-play noise-label correction framework for unsupervised domain adaptation person re-identification.
- Author
-
Tian, Qing and Du, Xiaoxin
- Subjects
- *
ARTIFICIAL neural networks - Abstract
Unsupervised domain adaptation person re-identification (UDA ReID) aims at leveraging knowledge from the source domain to help perform ReID in the unlabeled target domain. Most of existing investigations usually assign target instances identification labels through clustering them into the source person identification patterns. Unfortunately, inaccurate labels are frequently generated in such clustering setting, which undesirably deteriorates the accuracy of UDA ReID. Although a variety of noise label correcting works have been published to attempt addressing these misclustered labels, their effectiveness critically depends on the quality of the generated clustering representations, especially their differentiability, tending to result in limited efficacy. In this work, we design a desirable plug-and-play noise-label correction (PP-NLC) framework to efficiently correct the predicted target domain noise labels. Specifically, in PP-NLC we construct a noise label corrector (NLCr) and treat the predicted target domain noise labels as its probabilistic label variables (PLVs). Consequently, these noise labels are implicitly and automatically corrected through explicitly updating PLVs in the back-propagation process of NLCr, instead of conducting corrections directly on the generated noise pseudo-labels. Notably, the proposed PP-NLC framework enjoys desirable universality and can be deployed to existing UDA ReID approaches. Comprehensive experiments and analyses show the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF