Image-attribute reciprocally guided attention network for pedestrian attribute recognition

Abstract

Pedestrian attribute recognition is a fundamental yet challenging task in intelligent surveillance. Previous work focus on either discovering the discriminative visual features, or exploring the semantic dependencies among the attributes. However, the learning of comprehensive features by exploiting the relationship between the visual appearances and the attributes are not fully considered. To this end, we propose an image-attribute reciprocal guidance representation method, which investigates image-guided feature and attribute-guided feature to learn the comprehensive features for different attributes. To adaptively optimize the weight distributions of both features, we further develop a fusion attention mechanism. Besides, we present a focal cross-entropy loss to address the attribute imbalance problem. The whole framework is called Image-Attribute Reciprocally Guided Attention Network (IA2-Net), which is an end-to-end model. Extensive experiments on benchmark PETA and RAP datasets show that IA2-Net is very competitive with state-of-the art approaches.