Shape Translation And Editing Methods Based On Hidden Space Decoupling

With the development of 3D perception devices,people need more 3D data to meet the needs of games,AR/VR,3D movies and virtual worlds.At this time,3D point cloud data,as a data type that intuitively describes the real 3D world,has gradually become a hot spot in research.To address the lack of demand and supply for 3D models in a variety of fields,including gaming,AR/VR,and 3D film,traditional approaches rely on modeling tools and domain experts.But in recent years,with the rapid development of deep learning in computer graphics and computer vision tasks,recent techniques can enable them to create different sets of three-dimensional models without human intervention by using deep learning techniques and driven purely by data.Such techniques can be roughly divided into two categories: three-dimensional shape translation with reference to image translation tasks;and depth-based shape generation model-assisted 3D shape editing,which is different from traditional methods.Both 3D shape translation and 3D shape editing tasks are extremely challenging.The main work and innovation points of this paper are summarized as follows:In view of the current research status that the shape translation of point clouds can only achieve single-domain multi-style shape translation,this paper proposes a reversible explanatory network that uses a standardized flow model with reversible properties to decouple hidden variables in hidden space.The network uses a phased approach to building and training,which divides the network into reconstructed network and reversible interpretation network modules.Finally,by exchanging the different styles of the specified domain,the multi-domain and multi-style shape translation task of the point cloud is realized.Experimental results show that our method has certain advantages in the multi-domain and multi-style shape translation task of point cloud.In this paper,a lightweight shape editing method is proposed in view of the huge labor and time cost required by traditional methods in the current 3D data editing tasks,the huge amount of data required by recent technologies and the excessive parameters required for training.This method summarizes the characteristics of hidden space by proposing detailed research and analysis of the point cloud data representation and its corresponding hidden variable space,and then proposes that the core idea is to move the hidden variable to the hyperplane direction to obtain the hidden variable that crosses the shape of different semantic properties corresponding to the hyperplane.Experimental results show that our method has the properties of lightweight and easy to transplant to different spaces in the shape editing task of point clouds.