Zero-ReID: Towards Modality-Unified Generalizable Person Re-identification with Inter-Frame Identity Guidance

Generalizable person re-identification (ReID) focuses on learning identity-discriminative and domain-invariant features across multiple unseen scenarios, achieving significant progress on the canonical ReID. However, the generalization ability of visible-infrared person re-identification (VI-ReID) remains an open challenge, which has rarely been explored in the ReID community. To address this issue, we propose Zero-Shot ReID (Zero-ReID) framework, designed to tackle both generalizable canonical RGB and VI-ReID within a modality-unified model, which leverages large-scale unlabeled inter-frame data for contrastive learning. Zero-ReID intends to generalize to unseen real-world scenarios across both RGB and infrared modalities. Initially, we adopt the data synthesis techniques to translate RGB modality into missing infrared modalities. Then, Zero-ReID establishes identity correspondence priors between inter-frame visible images via bipartite graph matching. Finally, multiple priors-guided cross-modality correspondences are introduced to learn identity-discriminative and modality-shared features, while boosting performance on both modalities. Extensive experiments demonstrate that Zero-ReID achieves state-of-the-art zero-shot performance on 9 unseen public ReID datasets, with 87.9% Rank-1 on Market-1501, and 21.7% Rank-1 on SYSU-MM01.