Convolutional Attribute Mask with Two-step Attention for Fashion Image Retrieval

We propose a method to learn multiple latent spaces for attribute-specific fashion image retrieval. Our network learns multiple deep image features for a given set of fashion attributes through convolutional attribute masks and two-step attention. The masks promote our network to learn image features for a specific attribute. The two-step attention helps our network collect important spatial and channel information for the fashion attribute using spatial or channel attention. We visually show that our network correctly attend to essential regions for a given fashion attribute and learns well-distanced embeddings in latent spaces. We achieve state-of-the-art performance for attribute-specific fashion image retrieval.