3D Human Reconstruction Based On Deep Learning From Single-view Image

With the continuous development of network technology and hardware equipment,3D reconstruction technology has been applied in all walks of life,and has actually changed people’s lives.At the same time,the concept of the metaverse that has been born in recent years has gradually developed into one of the hottest topics in today’s society.As one of the core technologies of the metaverse,human body 3D reconstruction technology can provide users with better immersive Interactive experience.With the rapid development of computer vision technology,higher requirements are put forward for the integrity of the three-dimensional reconstruction model of the human body and the simplicity of the reconstruction method.Therefore,how to efficiently and easily reconstruct a complete and detailed 3D model of the human body has become a hot research issue.Against the above background,this topic focuses on the subject of 3D human body reconstruction based on deep learning,and mainly studies how to use its incomplete information to reconstruct a complete and detailed 3D human body model for single-view images.The main research contents are summarized as follows:(1)Parametric human body model reconstruction based on HMR algorithm.In order to In order to avoid the problems of incomplete reconstruction due to lack of information in the existing single-view image reconstruction methods,this paper proposes a method of introducing a parametric model of the human body to obtain complete 3D information of the human body,and uses the 3D regression module in the HMR algorithm to estimate the parametric human body model parameters,and then obtain the 3D skeleton information of the target person through the reconstructed SMPL model to provide prior information for subsequent experiments.(2)Normal graph estimation based on generative adversarial networks.In view of the lack of details in the current 3D human body reconstruction method based on implicit functions,this paper proposes to obtain normal information by estimating the normal maps of the front and back of the target person.Use the pix2 pix HD network as the normal map estimation network,and train two kinds of networks for the front and back normal map estimation networks respectively.Experiments show that the trained network can completely reconstruct the normal map corresponding to the target person,and provide normal prior information for subsequent reconstruction work.(3)Skeleton-based structural prior feature acquisition.The existing 3D human body reconstruction methods based on implicit functions focus more on the human body surface,and are prone to problems such as reconstruction results not conforming to the real human body structure.In this regard,this paper proposes a skeleton-based structure prior method,which uses the three-dimensional skeleton information of the human body to guide the network to learn the connectivity between the human skeleton and bones.Experiments show that adding this prior information effectively improves the authenticity of the reconstruction model.(4)3D reconstruction of the human body based on geometric priors.The 3D reconstruction method based on implicit function is optimized,adding the above-mentioned additional prior information as input,using the implicit function to obtain an occupancy field,and finally using the Marching Cubes algorithm to interpolate to obtain the final human body representation.The experimental results show that the optimization algorithm proposed in this paper can realize the three-dimensional reconstruction of the human body through single-view images.Compared with the existing methods on the same data set,the method in this paper performs better in the integrity and texture details of the reconstructed model...