Relation-Aware Attribute Network for Fine-Grained Clothing Recognition

In recent years, clothing attribute recognition made significant progress in the development of fine-grained clothing datasets. However, most existing methods treat some related attributes as different categories for attribute recognition in these fine-grained datasets, which ignores the intrinsic relations between clothing attributes. To describe the relations between clothing attributes and quantify the influence of attribute relation on attribute recognition tasks, we propose a novel Relation-Aware Attribute Network (RAAN). The relations between clothing attributes can be characterized by the Relation Graph Attention Network (RGAT) constructed for each attribute. Moreover, with the combination of visual features and relational features of attribute values, the influence of attribute relations on the attribute recognition task can be quantified. Extensive experiments show the effectiveness of RAAN in clothing attribute recognition.