Mesh Deformation Study Based On Interpolation Loop Subdivision

Deep learning has achieved rapid development in the field of image processing.And objects in the real world are represented by three-dimensional models.Applying deep learning to three-dimensional models has become one of the research hotspots of researchers.Grid is a representation form of 3D model,and the deformation of grid is an important research topic in computer graphics.At present,the main mesh deformation algorithm is deformation,and few deep learning methods are introduced.The computation speed of deep learning is fast,and the ability of predicting the target grid is strong.In deep learning,Loop subdivision is used in most mesh deformation methods.The method is based on the triangular mesh to do the approximate subdivision model,the subdivision mesh surface is smooth.However,using this subdivision algorithm to achieve mesh deformation will lead to the deformation of the mesh surface is not beautiful,easy to produce broken corners,holes.To solve these problems,this dissertation focuses on the subdivision method.And it designs two methods to study the mesh deformation algorithm,so as to improve the mesh deformation effect.1.In the Pytorch3 D mesh deformation algorithm based on Loop subdivision,the deformation result mesh is prone to produce broken corners and sparse surfaces.And it cannot fit the sharp features of the target mesh well.This dissertation proposes a Pytorch3 D mesh deformation algorithm based on interpolation Loop subdivision,referred to as ILP3 D algorithm.In this algorithm,the grid subdivided by interpolation Loop of regular dodecahedron is used as the source grid.And it selects the target mesh,then it samples the same point cloud on the surface of the source mesh and the target mesh at the same time.It calculates the Chamfer distance the two sets of point clouds.Then the source mesh offsets its vertices to complete the mesh deformation,gradually fitting the target mesh.Experimental results show that the modified algorithm is not easy to produce broken angles and sparse surfaces.At the same time,the Hausdorff distance decreases in most cases,which can better fit the detail features of the target mesh,and the visualization effect of the deformation result mesh is better.Moreover,the visualization effect of the deformation result mesh is better.2.Mesh-based 3D reconstruction requires the use of a source mesh to gradually transform into a target object.Aiming at the fact that the mesh of Pytorch3 D 3D reconstruction is prone to produce broken corners,holes and slender triangles,this dissertation proposes a Pytorch3 D 3D reconstruction algorithm based on interpolation loop subdivision,referred to as ILP3 DR algorithm.The algorithm first samples the textured mesh from 20 viewpoints to obtain 20 images.Next,it uses the Res Net-50 network to extract image features.And the 3D features are extracted by the 3D convolution module.It uses the regular dodecahedron subdivided by the interpolation Loop as the source mesh.Then the extracted features are projected onto the grid vertices.Finally,the source grid is deformed by the graph convolution network to obtain the reconstructed 3D grid.The experimental results show that the mesh reconstructed by the modified algorithm is not easy to produce broken corners,voids and slender triangles.Of course,the total loss value and Hausdorff distance of the mesh are significantly reduced.Not only can it make the reconstructed mesh shape is more similar to the target mesh,but also it make the visualization effect of the deformation result mesh is better.