Body Talk: Crowdshaping Realistic 3D Avatars with Words

摘要

Realistic, metrically accurate, 3D human avatars are useful forrngames, shopping, virtual reality, and health applications. Suchrnavatars are not in wide use because solutions for creating them fromrnhigh-end scanners, low-cost range cameras, and tailoring measurementsrnall have limitations. Here we propose a simple solution andrnshow that it is surprisingly accurate. We use crowdsourcing to generaternattribute ratings of 3D body shapes corresponding to standardrnlinguistic descriptions of 3D shape. We then learn a linear functionrnrelating these ratings to 3D human shape parameters. Givenrnan image of a new body, we again turn to the crowd for ratings ofrnthe body shape. The collection of linguistic ratings of a photographrnprovides remarkably strong constraints on the metric 3D shape. Werncall the process crowdshaping and show that our Body Talk systemrnproduces shapes that are perceptually indistinguishable from bodiesrncreated from high-resolution scans and that the metric accuracyrnis sufficient for many tasks. This makes body “scanning” practicalrnwithout a scanner, opening up new applications including databasernsearch, visualization, and extracting avatars from books.