[4L2-GS-4] Web intelligence

[4L2-GS-4-04] Applications and Further Developments of a Fashion Intelligence System for Fashion-Specific Ambiguous Expressions Interpretation

〇Ryotaro Shimizu^1,2, Yuki Saito², Masayuki Goto¹ (1. Waseda University, 2. ZOZO Research)

Keywords:vision and language, fashion interpretation, visual-semantic embedding, fashion intelligence, explainable artificial intelligence

Ambiguity in consumers' preferences and interpretations of fashion leads to hindering their deep interest in fashion by making it difficult for them to try new fashion genres. For this challenge, the authors propose a new technology called "Fashion Intelligence System," based on a vision and language model that automatically learns and interprets fashion images. The system utilizes a massive volume of full-body outfit images with attached tags, which often contain ambiguous expressions, to clarify the relationship between outfits and various expressions and to obtain answers to ambiguous and difficult-to-answer questions from users. Multiple models have been proposed to realize this system, but their applicability varies greatly depending on features such as how components are mapped into the projective space. In this study, we introduce an overview of the fashion intelligence system, present concrete examples of its applications based on experiment results, and discuss its prospects for further development.