Cross-modal Pedestrian Re-Identification Based On Fine-Grained Information And Non-Local Attention Mechanism

Abstract

Cross-modal pedestrian re-identification is used to match pedestrian images in different modes (infrared mode and visible light mode).For example,there is a pedestrian picture collected in the visible light mode during the day.The goal is to determine whether the same pedestrian has appeared in the picture collected by the infrared camera at night,and vice versa.Cross-modal pedestrian re-identification mainly solves the problem of pedestrian re-identification in weak light and night.This paper aims to improve the feature representation ability of the network and design appropriate loss functions to improve the similarity of the same pedestrian in two modes.In order to improve the performance of feature representation of feature learning module,a deep neural network structure combining triple loss function and nonlocal mechanism is introduced.On the other hand,it improves the cross-modal similarity within the class.In order to extract the fine-grained information of pedestrians,the pedestrian image is horizontally divided into four part,fine-grained feature learning further enhances the discriminant of network feature representation from both local and global aspects.The method of data enhancement by random erasure is added to improve the generalization of the test.The network achieves good performance on SYSU-MM01 and RegDB data sets.