Adaptive Style Transfer Learning for Generalizable Person Re-identification

Domain Generalization person re-identification (DG-ReID) is an emerging field with the ability to excel in unseen domains and the promise for fast deployment. However, its training phase, lacking access to target data, poses challenges. Simply extracting features through the backbone network and doing binary classification tasks, the DG-ReID method often faces the risk of overfitting, thereby impairing their ability to generalize effectively in new domains. Within a significant number of DG-ReID studies, an approach of data reduction is employed to mitigate the overfitting issue. This typically incorporates the use of Instance Normalization (IN), a method intended to cultivate domain-agnostic features. However, in the process, it may inadvertently eliminate basic person feature information. Consequently, we propose the inception of a novel Style Transfer Learning (STL) framework. This innovative scheme capitalizes on data augmentation—Mainly through style transformation techniques to generate new styles using known domain styles in the training set. At the same time, to prevent the generated new style from deviating too much, we also use a logic pairing model for correction to stabilize the generated new style. A powerful Style Transfer Learning framework stimulates the generation of a vast array of stable feature styles within the data whilst significantly enhancing the model’s generalization performance, thereby preventing the model overfitting from with specific domain styles. A wealth of experimental data validates our proposed strategy’s efficacy when applied to numerous DG-ReID datasets.