摘要
基于计算机视觉的虚拟试衣(virtual try-on,VITON)技术是指将试穿服装按照模特图像特征进行扭曲并合成到模特图像中,以替换原有服装部分。当前的虚拟试衣技术主要存在两个问题:保留模特图像头部、下装和背景等原有特征不足;扭曲后的试穿服装与模特图像匹配度不高。针对这两个问题,提出一种原有特征保持虚拟试衣网络(original feature preserving virtual try-on network,OFP-VTON),由语义分割图生成、试穿服装扭曲和试穿图像合成三部分组成。在试穿服装扭曲阶段通过使网络学习模特图像中所穿服装的扭曲映射,以更好地约束试穿服装扭曲;在试穿图像合成阶段提取并保留模特图像原有特征,并引入感受野模块(receptive field block,RFB)以尽可能保留试穿服装特征。在公开的VITON数据集上的定性与定量实验表明,OFP-VTON能更好地保留原有特征,扭曲后的试穿服装与模特图像匹配度高。
Computer vision-based virtual try-on(VITON)technology refers to warping and composing the try-on clothing according to the model image features into the model image to replace the original clothing parts.Current VITON methods have two main challenges:insufficient preservation of original features such as the head,bottom,and background of the model image;poor matching of the warped try-on clothing to the model image.To solve these two problems,an original feature preserving virtual try-on network(OFP-VTON)is proposed,which consists of semantic segmentation map generation,try-on clothing warping,and try-on image synthesis.In the try-on clothing warping phase,the network learns the mapping of warping of the clothing worn in the model image to better constrain the try-on warping.In the try-on image synthesis phase,the original features of the model image are extracted and preserved,and a receptive field block(RFB)is introduced to preserve the features of try-on clothing as much as possible.Qualitative and quantitative experiments on the publicly available VITON dataset show that the proposed OFP-VTON better preserves the original features and that the warped try-on clothing matches the model images better than the baseline method.
作者
王昭阳
陶然
卢海伦
WANG Zhaoyang;TAO Ran;LU Hailun(College of Computer Science and Technology,Donghua University,Shanghai 201620,China)
基金
National Key Research and Development Program of China(No.2020YFB1707700)
Fundamental Research Funds for the Central Universities,China(No.20D111201)。
