Cross-modality person re-identification, also known as RGB-infrared person re-identification, is a person recognition and retrieval technology between visible and infrared images. As an essential technologies of video intelligent surveillance system, the technology is urgently needed in security monitoring and suspect tracking. The modality discrepancy and intra-class variations in cross-modality person re-identification make it difficult to extract effective discriminable features of person identity. There are significantly fewer effective discriminable features in the shared feature space caused by shared network parameters. In order to solve this problem, an attention-guided cross-modality person re-identification is designed in this paper. The attention module uses the attention mechanism to focus on the salient features, so as to guide the network to put more parameters on the learning of salient features, so that the network can learn more discriminable, features from salient features. A comparative experiment on two standard data sets of cross-modality person re-identification verifies the effectiveness and advancement of the proposed method. For example, the proposed method improves rank-1 by 17.79% on SYSU-MM01 and 16.55% on mAP.
|