UVTON: UV Mapping to Consider the 3D Structure of a Human in Image-Based Virtual Try-On Network
UVTON: UV Mapping to Consider the 3D Structure of a Human in Image-Based Virtual Try-On Network
复制标题
DOI：
10.1109/iccvw.2019.00375
复制
发表时间：
2019-10
期刊：
影响因子：
--
通讯作者：
Shizuma Kubo;Yusuke Iwasawa;Masahiro Suzuki;Y. Matsuo
中科院分区：
文献类型：
--
作者：
Shizuma Kubo;Yusuke Iwasawa;Masahiro Suzuki;Y. Matsuo
随着在线服装购物需求的不断增加，基于图像的虚拟试穿是一个备受关注的研究领域。迄今为止提出的方法主要集中在如何在保留服装细节的同时生成装扮图像。然而，图像中模特的姿势仅限于直立姿势，其他姿势往往效果不佳。在这项研究中，基于一种利用 UV 映射考虑人体 3D 结构的生成对抗网络 (GAN)，我们提出了一种称为 UV 试穿网络 (UVTON) 的新型虚拟试穿方法。我们使用 DensePose 为 2D 图像的每个像素点估计与人体模型 3D 表面相对应的点，并将估计的信息合并到我们的模型中。因此可以改变处于各种姿势的用户的衣服。我们提出的方法使用 UV 映射和其他两个模块。一个模块生成要在映射中使用的部件，另一个模块细化图像并生成更真实的图像。基于与现有方法的定性和定量比较，我们通过实验证明我们的方法在各种姿势下都能取得更好的结果。
Image-based virtual try-on is an area of research that is attracting attention as the demand for online apparel shopping continues to increase. The methods proposed thus far have focused on how to generate a dress-up image while preserving the clothing details. However, the posture of the model in the image is limited to an upright position, and other positions frequently do not work well. In this study, based on a kind of generative adversarial network (GAN) that utilizes UV mapping to consider the 3D structure of the human body, we propose a novel virtual try-on method called a UV Try-On Network (UVTON). We use a DensePose to estimate a point corresponding to the 3D surface of a human model for each pixel point of a 2D image and incorporate the estimated information into our model. It is thus possible to change the clothes of users holding various postures. Our proposed method uses UV mapping and two other modules. One module generates parts to be used in the mapping, and the other refines the image and produces a more realistic image. Based on both qualitative and quantitative comparison with existing methods, we experimentally demonstrated that our method achieved better results with various postures.