Visible-infrared person re-identification model based on feature consistency and modal indistinguishability

Visible-infrared person re-identification (VI-ReID) is used to search person images across cameras under different modalities, which can address the limitation of visible-based ReID in dark environments. Intra-class discrepancy and feature-expression discrepancy caused by modal changes are two major problems in VI-ReID. To address these problems, a VI-ReID model based on feature consistency and modal indistinguishability is proposed. Specifically, image features of the two modalities are obtained through a one-stream network. Aiming at the problem of intra-class discrepancy, the class-level central consistency loss is developed, which makes the two modal feature centroids of the same identity close to their class centroids. To reduce the discrepancy of feature expression, a modal adversarial learning strategy is designed to distinguish whether the features are consistent with the modality attributes. The aim is that the generator generates the feature which the discriminator cannot distinguish its modality, thereby shortening the modality discrepancy and improving the modal indistinguishability. The generator network is optimized by the cross-entropy loss, the triplet loss, the proposed central consistency loss and the modal adversarial loss. The discriminator network is optimized by the modal adversarial loss. Experiments on the two datasets SYSU-MM01 and RegDB demonstrate that our method has better VI-ReID performance. The code has been released in .