For unsupervised VI-ReID, prior works center efforts on developing modality-shared features for intra-modality clustering and inter-modality matching, yielding promising results. However, large modality gaps hinder cross-modal co-occurrence mining, especially without cross-modal identity annotations. This often results in identity-ambiguous clusters and poorly aligned pseudo-labels, as shared features alone lack discriminative power and fail to model modality-specific variations. Given this, why not leverage modality-specific features to refine both clustering and matching in specific-feature spaces, as they can capture complementary identity cues absent in shared features? To this end, we deliver a novel Shared-Specific Synergy Network (S3Net), which embraces a Full-modal Clustering Representation Learning (FCRL) and a Shared-Specific Pseudo-label Synergy (SSPS) modules. In particular, FCRL first performs a “Decomposition-then-Compensation” tactic to furnish full-modal depictions (encompassing visible and infrared attributes) for unimodal pedestrians, enabling triple intra-modality clustering in three feature spaces: one modality-shared and two modality-specific ones. On that basis, SSPS further harvests the complementary inter-modality matching between the decomposed and compensated specific features. This can calibrate naïve cross-modality associations originally biased toward shared features in a multi-label space, eliminating the disturbances from noisy matches. Experimental results demonstrate that explicitly modeling modality-specific information can noticeably suppress noisy correspondences and tighten cluster purity. This work also highlights, for the first time, the potential of modality-specific cues for unsupervised VI-ReID.
