Deep Fashion Analysis with Feature Map Upsampling and Landmark-Driven Attention

Jingyuan Liu, Hong Lu

Abstract: In this paper, we propose an attentive fashion network to address three problems of fashion analysis, namely landmark localization, category classification and attribute prediction. By utilizing a landmark prediction branch with upsampling network structure, we boost the accuracy of fashion landmark localization. With the aid of the predicted landmarks, a landmark-driven attention mechanism is proposed to help improve the precision of fashion category classification and attribute prediction. Experimental results show that our approach outperforms the state-of-the-arts on the DeepFashion dataset.