Abstract
Garment sewing pattern represents the intrinsic rest shape of a garment, andis the core for many applications like fashion design, virtual try-on, anddigital avatars. In this work, we explore the challenging problem of recoveringgarment sewing patterns from daily photos for augmenting these applications. Tosolve the problem, we first synthesize a versatile dataset, named SewFactory,which consists of around 1M images and ground-truth sewing patterns for modeltraining and quantitative evaluation. SewFactory covers a wide range of humanposes, body shapes, and sewing patterns, and possesses realistic appearancesthanks to the proposed human texture synthesis network. Then, we propose atwo-level Transformer network called Sewformer, which significantly improvesthe sewing pattern prediction performance. Extensive experiments demonstratethat the proposed framework is effective in recovering sewing patterns and wellgeneralizes to casually-taken human photos. Code, dataset, and pre-trainedmodels are available at: https://sewformer.github.io.