General domain generalized person re-identification via mixing generic and specialized features

Person re-identification (ReID) has achieved promising performance under the assumption of a closed environment. However, in such an environment, solving a new task requires accessing the training data of all new tasks to retrain the model. Due to concerns about data privacy and timeliness, the practicality of these methods is limited. In this paper, we explore the general domain generalization problem for ReID, which aims to train a model that performs well on unseen domains in single-domain, joint multi-domain, or continuous multi-domain scenarios. To address this issue, we adopt pre-trained models and knowledge distillation to learn domain-invariant discriminative features. Additionally, we employ a sample generation algorithm based on extrapolation to learn the domain-specific discriminative features for important tasks and complement them with domain-general discriminative features. Finally, we conducted performance evaluations under three domain generalization protocols, and our method achieved promising results in all cases.