DMANet: Dual-modality alignment network for visible-infrared person re-identification

Xu Cheng, Shuya Deng, Hao Yu, Guoying Zhao

Abstract: Highlights•We propose a novel dual-modality alignment network (DMANet) to solve the problem of inter- and intra- modality alignments.•We develop an effective multi-granularity features mutual learning (MGFML) module to relieve modality discrepancy. Further, the domain maximum mean discrepancy loss and mutual learning loss are utilized to enhance the identity-aware ability.•We present an effective inter- and intra- modality alignment (IIMA) module to explore the interactions of inter- and intra- modalities based on the global self-attention mechanism, which solves the problem of insufficient global feature discrimination.•Extensive experimental results have verified the effectiveness of the proposed method. In addition, we have also verified the strong generalization of our approach on two corrupted cross-modality datasets.