| 3D clothed human reconstruction has a wide range of applications in military simulation,VR/AR content creation,image/video editing,telepresence,and virtual dressing.With the concept of the metaverse,avatars have become an important part of this 3D online social space.As a result,fine-grained 3D clothed human reconstruction technology has gained more attention.At present,3D clothed human reconstruction faces problems such as low reconstruction accuracy,a simple human pose,and limited input perspective.In this paper,we address the problems related to 3D clothed human reconstruction from two aspects: fine-grained human reconstruction and garment reconstruction based on the human model.For fine-grained human reconstruction,the pixel-aligned implicit function(PIFu)representation is more memory efficient than several existing 3D body representations,can recover shapes at arbitrary resolutions,captures finer local shapes,and solves the feature ambiguity problem inherent in single-view reconstruction.However,it does not take into account global geometric consistency,which may lead to unnatural body shapes being reconstructed.Inspired by the self-attentive Transformer-based model,we propose Trans PIFu,a single-view body reconstruction method based on Transformer and pixel-aligned implicit functions.In Trans PIFu,we lift the input 2D image features into 3D geometric features and divide them into several 3D feature block sequences.We constrain the global shape by exploring the geometric relationships between the 3D sequences.Compared with existing methods,the reconstruction results of Trans PIFu have less shape distortion,better surface details,and are more efficient than the two-stage reconstruction methods.The experimental results on the Deep Human dataset and the BUFF dataset demonstrate that Trans PIFu has better reconstruction results both globally and locally.For garment reconstruction,the reconstruction accuracy of existing garment reconstruction methods is low,and most of the garment styles that can be reconstructed are for tight-fitting models.Inspired by traditional clothing design methods,we propose a 3D reconstruction method GFLNet based on garment feature lines.Based on using SMPL for human reconstruction,we propose the key 3D feature lines for reconstructing garments and establish the geometric relationship between feature lines for the geometric structure of the standard garment.Compared with the existing methods,GFLNet can make the dimensions of the garment mesh independent of the human mesh,support the personalization of the garment,significantly improve the expressiveness of the garment mesh model and the accuracy of the geometry,and also support more types of garment topologies.The experimental results on the improved synthetic dataset and BUFF dataset prove that GFLNet has a better reconstruction effect,especially in the local structure of garments(such as cuffs,necklines,etc.),to achieve finer representation.To verify the feasibility and application scenarios of the two algorithms,we integrate the human reconstruction and garment reconstruction algorithms into a "virtual clothed body" generation application.By inputting a single clothed body image,the user can observe the 3D body model and garment model that correspond to the target image.Through experiments and system validation,the effectiveness of the method is confirmed,and the method is proved to be an important guide for the development of smart life. |
