关键词: 虚拟试衣, 三维人体建模, 二维点云, 黑白二值图, 生成对抗网络, 数据增强

Abstract:

In recent years, the method of 3D object modeling based on 2D images has developed rapidly. But in terms of the human body modeling, since the two-dimensional body image captured by the camera contains a lot of texture information such as clothes and hair. However, related applications such as virtual fitting need to remove texture information such as clothing wrinkles on human body surface. Meanwhile, considering that naked data collection violates the privacy of users, so this paper puts forward a new kind of modeling method, based on two-dimensional point cloud image to 3D human body model reconstruction method. Different from the collection of 2D image data set by camera and other auxiliary equipment, this paper directly draws 3D model samples from 3D human point cloud model library by vertex model to obtain point cloud rendering sample. The main work of this paper is to establish a data set consisting of a two-dimensional point cloud and the corresponding black and white binary image of the human body, and train a generative adversarial networks model generated by the former to generate the latter. By generative adversarial networks, the obtained two-dimensional point cloud image is transformed into the corresponding black and white binary image of human body. The black and white binary graph learned from the generative adversarial networks is input into a trained convolutional neural network, which is used to evaluate the construction effect of the 2D image to the 3D human modeling. Considering that it is a challenging problem to reconstruct a complete 3D human mesh model from an incomplete 3D point cloud, the method enables the algorithm to process an incomplete 2D point cloud image by simulating the damaged and incomplete state of 2D point cloud. A large number of experimental results show that the 3D human modeling reconstructed by this method can effectively achieve the sense of visual reality. In order to make a quantitative analysis of the accuracy after reconstruction, a representative waist circumference of human features is selected as the error evaluation. At the same time, in order to increase the diversity of human body shape in the 3D mannequin database, a convenient data enhancement technology for 3D human body is introduced. Experimental results show that the method proposed can quickly create the corresponding digital mannequin by inputting only a two-dimensional point cloud image.

Key words: virtual fitting, 3D human body modeling, 2D point cloud, black and white binary graph, generative adversarial network, data enhancement