Research On 3D Reconstruction Technology Of Garments Based On Single Image

How to reconstruct the geometry of the object from a single image is a challenging problem in the field of computer vision.3D reconstruction of fashion clothing with this technology is a hot topic in recent academic community.At present,existing methods based on physical modeling,silhouette,single view/multiple view geometry reconstruction,depth image,deep learning technology methods focus on the task such as 3D reconstruction of the human body.However,these works expressed the human body and clothing in a unified grid,and the reconstruction results lost a lot of details of the clothing,such as wrinkles and textures.Therefore,in order to solve this problem,this paper proposes Fashion3 D,the largest real 3D data set of clothing in the academia,and an end-to-end 3D reconstruction benchmark algorithm for clothing based on a single image,3DFashion Net.Centering on this theme,the main research work of this paper is as follows:(1)For the 3D reconstruction technology of single image,because the problem itself is quite different from the multi-view geometric reconstruction algorithm,and it is an ill-posed problem with non-closed solution space.The expressions of 3D data include voxel,point cloud,mesh(parameterized mesh),multi-view image and so on.In this paper,the theoretical principle of graph convolutional network is analyzed in detail,and the mapping relation between 2D image and 3D mesh model is effectively solved by using the graph convolutional network to deform the parametric template mesh.(2)Aiming at the problem of 3D reconstruction of clothing,this paper uses the idea of parametric modeling of SMPL body and adopts the method of partial cutting and deformation SMPL to conduct parametric modeling of clothing surface.A new concept of feature line fitting is proposed,which can complete the generation of clothing mesh with high quality and make the optimal approximation to the geometric details of clothing in the input image.Based on these innovative points,this paper proposes a multi-stage 3D reconstruction baseline neural network,3DFashion Net,which can directly reconstruct the clothing model from a single image.(3)At present,the work based on deep learning in academia is carried out in a data-driven way.The data sets that can be directly used by algorithms and have detailed annotation information are the most important resources in the deep learning work,without data set,the algorithm can not do the corresponding work.so this paper produces the largest 3D dataset of real clothing in the academia,which contains 563 pieces of clothing,10 categories,and 2078 dense point cloud models.Each model has corresponding detailed labeling information,including corresponding multi-view images,costume pose parameters and feature line labeling.This has made a certain basic contribution to the future academic work related to the theme of clothing.In summary,combined with the data set and algorithm proposed in this paper,the clothing part of a single image can be reconstructed into a high-quality 3D mesh model efficiently,which also has a good effect compared with other similar algorithms.