Fine-Grained Fashion Similarity Prediction by Attribute-Specific Embedding Learning

This paper strives to predict fine-grained fashion similarity. In this similarity paradigm, one should pay more attention the in terms of a specific design/attribute between items. For example, whether collar designs two clothes are similar. It has potential value many related applications, such as copyright protection. To end, we propose an Attribute-Specific Embedding Network (ASEN) jointly learn multiple attribute-specific embeddings, thus measure corresponding space. The proposed ASEN is comprised global branch and local branch. takes whole image input extract features from perspective, while zoomed-in region-of-interest (RoI) w.r.t. specified attribute able features. As different perspectives, they complementary each other. Additionally, branch, modules, i.e., Attribute-aware Spatial Attention Channel Attention, integrated make be locate regions capture essential patterns under guidance attribute, learned embeddings better reflect Extensive experiments on three fashion-related datasets, FashionAI, DARN, DeepFashion, show effectiveness for prediction its reranking. Code data available at https://github.com/maryeon/asenpp .