Global-Local Mutual Guided Learning for Person Re-identification

Abstract

Person Re-Identification (Re-ID) plays a significant role in intelligent surveillance systems. Existing popular methods mainly focus on locating regions with specific pre-defined semantics to learn local representations, where pedestrian part-level features are inefficient to fully utilize the global feature information. Besides that, some methods miss out semantic transition information of human body. In this paper, we propose an end-to-end feature learning strategy to get refined feature representations with global-local mutual guided learning. In order to explore global and local information, we design a Global-Local Mutual Guided Network (GLMG-Net). It contains two branches to learn global feature representations, and local feature representations, respectively. For mutual guided module, global features are combined with each local feature by the add-wise operation. In the training process, this module enables branches to guide each other. Comprehensive experiments conducted on the public datasets of Market-1501 and DukeMTMC-ReID indicate that our method outperforms state-of-the-art approaches in several cases. In particular, mean average precision (mAP) scores of our method on those benchmarks are 89.2% and 79.7%, respectively.