Runway to Realway: Visual Analysis of Fashion Sirion Vittayakorn1 Kota Yamaguchi2 Alexander C. Berg1 Tamara L. Berg1 University of North Carolina at Chapel Hill1 Tohoku University2 {sirionv,acberg,tlberg}@cs.unc.edu
[email protected] Abstract quantitative analysis of fashion both on the runway and in the real world, computationally, and at large scale, using Clothing and fashion are an integral part of our every- computer vision. To enable this study, we collect a large day lives. In this paper we present an approach to studying new dataset called the Runway Dataset, containing 348,598 fashion both on the runway and in more real-world settings, runway fashion photos, representing 9,328 fashion show computationally, and at large scale, using computer vision. collections over 15 years. Using this dataset we develop Our contributions include collecting a new runway dataset, computer vision techniques to measure the similarity be- designing features suitable for capturing outfit appearance, tween images of clothing outfits. Additionally, we make use collecting human judgments of outfit similarity, and learn- of the Paper Doll dataset [26] collected from the Chictopia ing similarity functions on the features to mimic those judg- social network, where people interested in fashion post and ments. We provide both intrinsic and extrinsic evaluations annotate pictures of their daily outfits. The combination of of our learned models to assess performance on outfit simi- these two datasets allows us to make initial steps toward an- larity prediction as well as season, year, and brand estima- alyzing how fashion trends transfer from runway collections tion.