Serialized Local Feature Representation Learning for Infrared-Visible Person Re-identification

Abstract

Infrared-Visible person re-identification is a kind of cross-modality person re-identification. The purpose of the task is that given a person image we need to find another image on the same person from gallery. The query images and gallery images are not only in RGB modality but in infrared modality as well. The cross-modality person ReID task can deal with the limitation of single modality because we usually can get images in more than one modality. In our work, we take advantage of both global feature and local feature. We use a dual-path structure to extract features from RGB images and infrared images respectively. Besides, we add the LSTM structure in each path to learn the serialized local features. The loss function consists of cross-entropy loss and hetero-center loss so that the model can bridge the cross-modality and intra-modality gaps to capture the modality-shared features and improve the cross-modality similarity. Finally, we do experiments on two datasets including SYSU-MM01 and RegDB, then compare with other methods in recent studies.