Abstract

We contribute a new dataset and a novel method for natural language basedfashion image retrieval. Unlike previous fashion datasets, we provide naturallanguage annotations to facilitate the training of interactive image retrievalsystems, as well as the commonly used attribute based labels. We propose anovel approach and empirically demonstrate that combining natural languagefeedback with visual attribute information results in superior user feedbackmodeling and retrieval performance relative to using either of thesemodalities. We believe that our dataset can encourage further work ondeveloping more natural and real-world applicable conversational shoppingassistants.

 

Quick Read (beta)