Adaptive confidence-driven learning and cross-modal hard sample mining for unsupervised visible-infrared person re-identification

Abstract

This research addresses the critical challenges in Cross-modal Visible-Infrared Person Re-ID (VI-ReID), including significant modal differences, lack of cross-modal correspondence, and pseudo-label noise accumulation. To mitigate these issues, we propose an innovative framework integrating an adaptive multidimensional enhanced clustering method and a confidence-driven dynamic label correction mechanism. Specifically, we design a dynamic clustering framework leveraging neighborhood consistency and intra-class distribution entropy to autonomously model data distributions. A confidence-driven dynamic label correction mechanism is introduced, employing multi-prototype similarity probability models to filter pseudo-label noise effectively. Moreover, a cross-modal feature alignment strategy based on optimal transport theory addresses many-to-many feature matching between visible and infrared modalities. Additionally, a Hard Sample Aware Contrastive Learning (HCL) strategy is implemented to enhance feature learning in complex data distributions through dynamic feature storage. Extensive experiments conducted on SYSU-MM01 and RegDB datasets, comprising 29,533 and 4120 image pairs, respectively, demonstrate the framework’s effectiveness. The proposed method achieves a 3.9% mAP improvement on average compared to state-of-the-art methods, highlighting its advantages in cross-modal feature alignment and pseudo-label optimization.