Infrared-visible cross-modal person re-identification via dual-attention collaborative learning

Abstract

Person re-identification is regarded as a retrieval task for searching the same person in different cameras, within which infrared-visible cross-modal re-identification (VI-ReID) is challenging because the inter-class distance is larger than the intra-class distance. In this paper, a dual-attention collaborative(DAC) learning method is proposed, which unites channel and spatial attentive deep features to obtain supplementary information for multiple classifiers via a cross-modal consistency constraint. A channel attention and part-wise spatial pooling are adopted for discriminative feature learning. A multiple-classifier strategy with a cross-modal consistency constraints is presented for the cross-modal identification. In this way complementary information among modality-sharable classifier and modality-specific classifier can be better utilized. The experimental results show that the proposed method distinctly outperforms the baseline method by a margin of 9.83% Rank-1 and 6.84% mAP on SYSU-MM01.