Identity-Assisted Network for Pedestrian Attribute Recognition

Pedestrian attribute recognition aims to accurately locate and extract high-level semantic attributes of pedestrians, and provide important support for pedestrian re-identification. The existing pedestrian attribute recognition method has achieved good recognition results in indoor and other single scences. However, it does not perform well in complex background with changes of illumination, viewing position and occlusion. In this work, we introduce pedestrian identity (ID) as auxiliary information for attribute recognition, and propose an identity-assisted pedestrian attribute recognition network (IA). The IA network uses ResNet-50 as the backbone network, removes the last fully connected layer, and then connects to a multi-branch network, which contains re-identification branch and attribute branch. The re-identification branch is used to extract pedestrian features, then use hierarchical clustering to generate pseudo IDs, which finally assists pedestrian attribute identification. Besides, we construct a quintuple loss function. Firstly, it constructs a intra-triple loss within an attribute. And then, it constructs an inter-triple loss between attributes according to the pseudo ID information to fully optimize the attribute space. The average accuracy mA of the IA model for all attributes on the PETA dataset exceeds 85%. Through comparative experiments, it can be proved that the IA model gets a bettrer performance on attribute recognition.