Position Attention-Guided Learning for Infrared-Visible Person Re-identification.

Infrared-Visible person re-identification is a challenging and fundamental task of associating the same person across visible and thermal cameras. Most of the studies focus on improving the global features to address the cross-modality issue, thus, some discriminative local and salient features are ignored by the deep models. A novel deep architecture named Dual-path Local Information Structure (DLIS) with Position Attention-guided Learning Module (PALM) is proposed to address the cross-modality issue for Infrared-Visible PReID task. The DLIS has two individual branches which contains a visible stream and an infrared stream to extract modality sharable features. The PALM can capture long-range dependencies and enhance the discriminative local feature representations to form the final feature descriptors. To supervise the network extracting discriminative features to shrink the margin of different modalities, the proposed model is conducted the joint supervision of cross-entropy loss function and hetero-center loss function. Compared with the recent studies, the proposed methods achieve the state-of-the-art on the two benchmark datasets including SYSU-MM01 and RegDB dataset.