Intermediate domain-based unsupervised domain adaptive visual-infrared cross-modal person re-identification

Visible-infrared cross-modal person re-identification aims at identifying the same person from images captured by non-overlapping cross-modal cameras, presenting a significant challenge in computer vision. Most existing methods only focus on the modality gap within a single domain, neglecting the domain gap between different domains, which negatively impacts performance when domains change. Therefore, we propose unsupervised domain adaptation approach. First, an intermediate domain generator is introduced to mix source and target domain representations. Second, a cross-modal clustering method is used to establish cross-modal correspondence in the target domain. Finally, weighted bipartite graph matching further aligns the visible and infrared modalities. To the best of our knowledge, this is the first attempt to apply unsupervised domain adaptation to visible-infrared cross-modal person re-identification. Extensive experimental on three benchmarks show that our method significantly improves generalization.