Unified pre-training with pseudo infrared images for visible-infrared person re-identification

Abstract

In the pre-training task of visible-infrared person re-identification(VI-ReID), two main challenges arise: i) Domain disparities. A significant domain gap exists between the ImageNet utilized in public pre-trained models and the specific person data in the VI-ReID task. ii) Insufficient sample. Due to the challenge of gathering cross-modal paired samples, there is currently a scarcity of large-scale datasets suitable for pretraininge. To address the aforementioned issues, we propose a new unified pre-training framework (UPPI). Firstly, we established a large-scale visible-pseudo infrared paired sample repository (UnitCP) based on the existing visible person dataset, encompassing nearly 170,000 sample pairs. Benefiting from this repository, not only are training samples significantly expanded, but pre-training on this foundation also effectively bridges the domain disparities. Simultaneously, to fully harness the potential of the repository, we devised an innovative feature fusion mechanism(CF\(^2\)) during pre-training. It leverages redundant features present in the paired images to steer the model towards cross-modal feature fusion. In addition, during fine-tuning, to adapt the model to datasets lacking paired images, we introduced a center contrast loss(C\(^2\)). This loss guides the model to prioritize cross-modal features with consistent identities. Extensive experimental results on two standard benchmarks (SYSU-MM01 and RegDB) demonstrate that the proposed UPPI performs favorably against state-of-the-art methods.