Piecewise Method For Large Scale Reconstruction And Optimization

Extracting 3D surface of high quality mesh from the point cloud is well studied in computer graphics.It involves approximation of scanned point cloud,filling of surface holes and remeshing of existing models.In this paper,we sample a small number of points as the vertices of our reconstructed mesh by applying Poisson disk sampling algorithm.Then the Voronoi tessellation is used to compute the connectivity of the vertices,which generates the initial mesh.In the process of Voronoi tessellation,the shortest path is taken as the distance function to reduce the effect caused by nonuniform sampling.It can be seen that the coarse mesh is hard to achieve our desired effect in terms of both geometry and topology.Thus,a proper optimization scheme is required to improve the quality of the initial mesh.The whole optimization process is represented by the vertex set V and the adjacency matrix B as a constrained q-norm optimization problem to find the vertex position and connectivity of the mesh that minimize an energy function composed of point-to-mesh metric and regularization.In the process of optimization,the normal information is optional,which is required in many existing surface reconstruction algorithms.As we know,normal estimation can be extremely sensitive to noisy data and sharp features.In the process of reconstruction,inaccurate normal information often leads to large approximation errors.In contrast,our method is able to handle sampling data with unreliable normal information.In our algorithm,the normal information is used to compute the orientation of the face,but the normal information of high accuracy is not a must.Since the algorithm proposed in this paper only needs the geometric information of the sampling points near the vertex to update the topology and geometric information of the initial mesh,the initial point cloud could be divided into multiple sub-blocks and the appropriate surface is reconstructed for each sub-block.The final surface of the whole scene can be generated by stitching all surfaces together.Experiments show that the surface produced in this way have a good performance in terms of efficiency and accuracy when dealing with large scale point cloud.