3D Object Reconstruction With Generative Adversarial Networks

With the rapid development of computer technique,the boundary between machine cognition and human perception becomes more and more blurred.Therefore,the single two-dimensional data pattern can no longer satisfy people's exploration of the threedimensional world.How to quickly and accurately restore the widely existing twodimensional information to the three-dimensional space has become a key research subject in the field of computer vision and even industry.Three dimensional reconstruction system,as an effective means to complete the mapping from twodimensional information to three-dimensional space,can make full use of the information of two-dimensional image features and projection contour to automatically build the corresponding three-dimensional model,aiming to solve the problem of low-dimensional information dissatisfaction and better perceive and interact with the object or space.3D reconstruction,especially 3D object voxel reconstruction,is a new topic in the field of computer vision.It refers to the 3D shape recovery of target object or scene by using multi-view or even single view.But it is different from the general visual problems,its task is to perceive the three-dimensional environment,rather than simple recognition,detection,etc.At present,the commonly used solution of deep learning is to use the complex prior probability problem to express the dimensional upgrading transformation of space.However,the existing 3D reconstruction algorithms based on deep learning mostly adopt the heuristic criterion of single voxel independence,which excessively relies on the prior data and restricts the object contour.With the rise of 3D generative adversarial network,it provides a new idea for 3D reconstruction.It uses adversary criteria instead of traditional heuristic criteria to implicitly capture the object structure and lay a foundation for unsupervised reconstruction.Therefore,in this paper,the 3D object reconstruction is taken as the research focus,and two kinds of problems of singleangle 3D voxel reconstruction and unsupervised 3D voxel reconstruction based on generative adversarial networks are mainly explored.The main work and contributions are summarized as follows:Firstly,the asymmetry problem between 2D and 3D data sets was improved and solved.Five kinds of CAD objects in 3D Shape Net Core were selected for threedimensional voxelization,and each 3D object was rendered and generated to 2D images and 2D projection sets for 3D voxel reconstruction experiments.Secondly,in order to improve the accuracy of the 3D voxel reconstruction and the stability of the network training,the reconstruction loss function based on the Chamfer distance and the discriminator gradient penalty method based on the Wasserstein distance are used on the basis of the 3D voxel reconstruction algorithm based on generative adversarial network.Hence,parameter optimization of each network module is realized.Experimental results show that the set of loss functions can effectively improve the accuracy of 3D voxel reconstruction with fixed resolution,and is better than other related methods.Finally,the 3D voxel reconstruction algorithms mostly rely on prior data,but the 3D data is difficult to collect and process,an unsupervised 3D voxel reconstruction algorithm based on generative adversarial networks was proposed.It makes full use of the perspective projection transformation principle to transform the voxel model from the three-dimensional coordinate system to the projection coordinates,and then uses the multi-view projection set to complete the adversarial criterion,so as to complete the 3D voxel reconstruction.Experimental results show that the proposed method is available and superior to similar unsupervised methods in accuracy.