Weakly Supervised Pedestrian Attribute Recognition with Attention in Latent Space

Abstract

Pedestrian attribute recognition is a key problem in intelligent surveillance. Relations between attributes and human body structures or relations among attributes are beneficial to attribute recognition, while the annotations are just image-level binary labels. In this work, we propose a novel pedestrian attribute recognition network that takes advantage of latent attribute localizations and local attribute relations to improve the performance of pedestrian attribute recognition. Our method generates latent attribute localization maps by weakly-supervised learning in latent attribute localization (LAL) module. These latent attribute localization maps are fed into the local attribute attention (LAA) module to extract local attributes, and local attributes are interacted with each other with the attention mechanism. Extensive experiments made on the publicly pedestrian attribute datasets of PETA and RAP show that our model outperforms previous methods.