XID: A protocol for evaluating identity consistency under domain shifts and reidentification method

Recent domain generalization methods for person reidentification aim to learn features that remain discriminative across domains to improve performance in unseen environments. Prior work has addressed domain shift through discrepancy reduction and alternative normalization strategies, while maintaining identity separability. However, these evaluations often rely on simplified settings with non-overlapping identities and limited visual diversity. To address this, we propose a new evaluation protocol that introduces identity transfer and significant appearance variation by constructing query and gallery sets from different domains. This setup enables a more realistic assessment of intra-class variation and inter-class discriminability. We further develop a learning framework specifically designed for this protocol, which enhances generalization by regulating achromatic information and projecting embeddings into a space that simulates unseen domains. The framework includes a self-regulating augmentation policy that adjusts transformation strength during training. Extensive experiments show consistent performance gains under both the proposed and standard protocols, establishing a more rigorous and practical benchmark.