Abstract:

In view of the vague effect of Conditional Analogy GAN（CAGAN） after changing clothes, poor performance when the length of the target clothes is inconsistent with that of the original clothes, too few details are preserved relative to the target clothes, a new virtual try-on method based on improved CAGAN is proposed. The improved CAGAN generates a preliminary result and gets the mask of the target clothes after wearing on the model. Then the target clothes are deformed with mask. The deformed clothes and the results of the first step are combined to obtain the final results. The experimental results show that the method solves the previous problems and achieves very good results.

Key words: Generative Adversarial Networks（GAN）, virtual try-on, conditional generation

摘要：

针对CAGAN（Conditional Analogy GAN）换衣后效果模糊，在目标衣服与原始衣服长短不一致时效果一般，相对目标衣服保留过少的细节等问题做了相关研究并对CAGAN进行了改进，提出了新的虚拟试衣方式。经过改进的CAGAN生成一个粗糙的结果，由该结果得到目标衣服穿在模特身上改变形状后的mask，接下来利用mask对目标衣服进行变形，综合变形的衣服和第一步的结果便得到最终的试衣图像。实验结果表明，该方法解决了前面存在的问题，而且取得了非常好的效果。

关键词: 生成式对抗网络, 虚拟试衣, 条件生成