Recently, person re-identification with domain generalization has focused on reducing the gap between source domains and unseen target domains. Meta Batch Instance Normalization (MetaBIN) is one of the most effective methods for addressing the domain gap problem. However, the lack of style variation from limited source domains still makes diversifying virtual simulations difficult. To alleviate this, an improved generalizable person re-identification method is proposed; when this method is combined with MixStyle and meta-leaning, it is called MetaMix. Initially, we represent the MixStyle layers of MetaBIN to create a diverse style during the meta-learning process. Moreover, fine-tuning a BIN module is shown to be applicable for pre-trained re-identification (ReID) models by batch normalization. Extensive experimental results show the effectiveness of the proposed method in a large-scale cross-domain scenario.

Person re-identification is a problem of identifying individuals across non-overlapping cameras. Although remarkable progress has been made in the re-identification problem, it is still a challenging problem due to appearance variations of the same person as well as other people of similar appearance. Some prior works solved the issues by separating features of positive samples from features of negative ones. However, the performances of existing models considerably depend on the characteristics and statistics of the samples used for training. Thus, we propose a novel framework named sampling independent robust feature representation network (SirNet) that learns disentangled feature embedding from randomly chosen samples. A carefully designed sampling independent maximum discrepancy loss is introduced to model samples of the same person as a cluster. As a result, the proposed framework can generate additional hard negatives/positives using the learned features, which results in better discriminability from other identities. Extensive experimental results on large-scale benchmark datasets verify that the proposed model is more effective than prior state-of-the-art models.