【摘要】 现有的跨模态行人重识别方法大多选择模态互转或直接将特征映射到共同特征空间，主要关注于缓解模态差异，而忽略同一模态下的类间差异，这不利于模型学习到具有辨别性的身份特征。论文提出一种基于改进的多头自注意力机制的非共享双流网络，分别提取可见光图像和红外图像的特征，并对特征做水平划分，使用局部特征来计算身份损失以及异质中心损失，利用多头自注意力机制根据像素本身的内容及其相对于相邻像素的空间位置来计算像素的注意力图，增强模型提取的信息容量，进而学习到同模态下不同类之间具有辨别性的身份特征。同时使用辅助模态来减缓可见光图像和红外图像在颜色信息上的差异，进一步减小模态差异。在SYSU-MM01和RegDB数据集上的mAP分别达到60.36%和76.10%，实验表明该方法的有效性。更多还原

【Abstract】 Most of the existing cross modal pedestrian re-identification methods choose modal transformation or directly map features to the common feature space,mainly focusing on mitigating modal differences,while neglecting the differences between classes under the same mode,which is not conducive to the model learning to identify identity features. This paper proposes a non shared dual stream network based on the improved multi head self attention mechanism,which extracts the features of visible and infrared images respectively,divides the features horizontally,uses local features to calculate the identity loss and heterocenter loss,uses the multi head self attention mechanism to calculate the pixel’s attention map according to the content of the pixel itself and its spatial position relative to adjacent pixels,and enhances the information capacity extracted from the model. Then the distinctive identity characteristics between different kinds under the same mode can be learned. At the same time,auxiliary modes are used to reduce the difference in color information between visible and infrared images,and further reduce the modal difference. The mAP on SYSU-MM01 and RegDB datasets reaches 60.36% and 76.10% respectively. The experiment shows that the method is effective.更多还原