Dynamic Uncertainty Learning with Noisy Correspondence for Text-Based Person Search

Text-Based Person Search TBPS seeks to identify individuals using textual descriptions. However, in real-world scenarios, noisy correspondences—under-correlated or even false-correlated image-text pairs—significantly degrade retrieval performance. Existing approaches often overemphasize negative samples, inadvertently amplifying this noise. To address these challenges, we propose the Dynamic Uncertainty with Noisy Correspondences DUNC framework, which introduces a novel Cross-modal Uncertainty Learning paradigm and a robust loss function, Dynamic Robust Loss DRL . Unlike conventional methods that rely on global representations, DUNC effectively mines fine-grained correspondences, improving alignment between text and image features. Furthermore, DUNC employs a Dirichlet distribution to model bidirectional evidence from cross-modal similarity, enabling the capture of alignment uncertainty and reducing the effects of large intra-class variations. Meanwhile, DRL adaptively selects and aggregates the most challenging negative samples, mitigating noise and capturing a richer distribution of negative samples. This design enhances robustness and representation quality, even in noisy environments. Extensive experiments on three benchmark datasets demonstrate that DUNC achieves strong resistance to noise and improves retrieval performance under both low- and high-noise conditions. The code is publicly available at url{https: github.com ASL-forever DUNC}.