Segment-Attention Augmented Dual-Contrastive Aggregation Learning for Unsupervised Visible-Infrared Person Re-Identification

In the unsupervised visible-infrared person re-identification (USL-VI-ReID) task, cross-modal retrieval of person images faces many challenges. This is not only due to the inherent modal differences between visible and infrared images, but also because the same person may appear in different backgrounds and different persons may appear in the same background in real scenes, and these complex background variations increase the difficulty of learning robust features, especially under unsupervised conditions with limited labeling data. To address these issues, we propose the Segment-Attention Augmented Dual-Contrastive Aggregation Learning Model (SA-ADCA), an advancement over the existing Augmented Dual-Contrastive Aggregation Learning (ADCA) framework. The SA-ADCA model introduces a novel Segment-Anything mechanism to suppress background pixels and enhance focus on the person regions, thereby improving feature discrimination. Furthermore, a multi-attention mechanism is incorporated within the shared feature layer, enabling the model to capture fine-grained cross-modal relationships and achieve superior alignment between visible and infrared features. These innovations improve the quality of the modality representation, the correlation between modal characteristics and optimize modality-agnostic characteristics. Extensive experiments show that the proposed framework significantly outperforms existing unsupervised methods and surpasses several supervised models. Specifically, the improvement over ADCA was 4.58% and 1.63% in the RegDB dataset and the SYSU-MM01 dataset.