Dynamic Mutual Alignment for Unsupervised Domain Adaptation Text-to-Image Person Re-Identification
IEEE Signal Processing Letters · 2026 · 33 · 1581-1585
DOI: 10.1109/lsp.2026.3682192 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Pengyu Gu und 2 weiteren Autor:innen. Die Arbeit nennt 32 Literaturangaben.
Abstract (Originalsprache)
Text-to-Image Person Re-identification (TIReID) enables pedestrian retrieval using natural language descriptions, but its performance degrades severely when deployed across domains due to distribution shifts and modality heterogeneity. While unsupervised domain adaptation (UDA) has been widely studied for image-based re-identification, directly extending existing UDA methods to TIReID is challenging, as visual and textual modalities exhibit distinct feature distributions. To address this, we explore Unsupervised Domain Adaptation Text-to-Image Person Re-identification (UDA-TIReID) to bridge the cross-modal discrepancy and cross-domain shift for robust pedestrian retrieval in unlabeled target domains. We propose Dynamic Mutual Alignment (DMA), a framework for UDA-TIReID that employs a decoupled, consensus-driven design to explicitly enforce cross-modal identity consistency during adaptation. Specifically, we design a Mutual Nearest-Neighbor Alignment (MNNA) method to mine reliable cross-modal correspondences in the target domain via neighborhood consensus and dynamic confidence thresholding. Experimental results on benchmark datasets demonstrate that DMA consistently outperforms state-of-the-art methods.
Zitieren
APA
Gu, P., Tao, Q., & Yu, Y. (2026). Dynamic Mutual Alignment for Unsupervised Domain Adaptation Text-to-Image Person Re-Identification. IEEE Signal Processing Letters, 33, 1581-1585. https://doi.org/10.1109/lsp.2026.3682192
Vancouver
Gu P, Tao Q, Yu Y. Dynamic Mutual Alignment for Unsupervised Domain Adaptation Text-to-Image Person Re-Identification. IEEE Signal Processing Letters. 2026;33:1581-1585. doi:10.1109/lsp.2026.3682192
BibTeX
@article{Gu2026,
author = {Gu, Pengyu and Tao, Qingchuan and Yu, Yanmei},
title = {Dynamic Mutual Alignment for Unsupervised Domain Adaptation Text-to-Image Person Re-Identification},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {1581--1585},
doi = {10.1109/lsp.2026.3682192}
}
EUVIMED