Research On 3D Human Body Reconstruction With Specific Posture Based On 2D Image

The society develop rapidly and the entertainments of people becomes more and more diversified,3D human body reconstruction is used in our daily life widely.In the fields of film and television animation,video game,virtual fitting and so on,3D human body reconstruction shows its important social value.In order to realize 3D human body reconstruction in a more convenient way,the demands for the reconstruction of 3D human body model from a single 2D image become more and more intense.Now,the popular way to reconstruct 3D human body model based on a single 2D image is to use parameterized standard 3D human body template.Firstly,the body shape parameters,posture parameters and other parameters of the target body are obtained from the 2D image,then the parameters obtained are used to drive the parameterized standard 3D human body template to make corresponding changes in the body shape and posture of the standard 3D human body template,Finally,a 3D human body model with the same shape and posture as the target human body in the 2D image is obtained.we studies the current methods of reconstruction of 3D human body model based on2 D image and solves some common problems in the current methods in the thesis.When the target human body presents T pose and A pose,natural bending of elbows and knees occurs in the 3D human body model reconstructed based on a single 2D image.In the thesis,improvements are made on the basis of SMPLify,and the improved method is named SMPLify-M,which is used to solve the problem of natural bending of elbows and knees of reconstructed 3D human body model when the target human body in 2D image shows extended posture.At present,there are several common ideas in the reconstruction of 3D human body model based on a single 2D image: neural network regression,iterative fitting optimization,and cyclic joint neural network regression and iterative fitting optimization.The reconstruction based on the idea of iterative fitting optimization is very slow.The idea based on neural network regression requires a dataset composed of a large number of paired 2D images and 3D data labels.In the initial stage,the convolutional neural network module of the reconstruction based on the idea of cyclic joint neural network regression and iterative fitting optimization can hardly provides reasonable initialization parameters for the iterative fitting optimization module,and the reconstruction process is slow.Due to lack of sufficient pairs of 2D images and 3D data labels,we design the idea of cascade joint neural network regression and iterative fitting optimization in the thesis.Based on the idea,we design the HMR-SMPLify-X to reconstruct 3D human body model based on a single 2D image.The experimental datas and images prove that the idea of cascaded joint neural network regression and iterative fitting optimization can achieve a faster reconstruction speed on the basis of guaranteeing the reconstruction quality in the thesis.