Spectral surface reconstruction Reporter: Lincong Fang 24th Sep, 2008
Point clouds
Surface reconstruction Unorganized Unoriented (no oriented normals) Non-uniform, sparse sampling Possibly with noise
Applications Computer Graphics Medical Imaging Computer-aided Design Solid Modeling
Approaches Delaunay\Voronoi based Implicit surfaces Deformable models Spectral Etc.
Approaches Delaunay\Voronoi based Unorganized, unoriented, non-uniform, noise
Approaches Implicit surfaces Unorganized, unoriented, non-uniform, noise
Approaches Deformable models Adrei Sharf, Thomas Lewiner, Ariel Shamir, Leif Kobbelt, Daniel Cohen–OR. Competing fronts for coarse–to–fine surface reconstruction. EG2006
Approaches Delaunay\Voronoi based Implicit surfaces Deformable models Spectral Etc. [1] R. Kolluri, J. Richard Shewchuk, J. F. O’Brien, Spectral surface reconstruction from noisy point clouds. SGP [2] P. Alliez, D. Cohen-Steiner, Y. Tong, M. Desbrun Voronoi-based variational reconstruction of unoriented point sets. SGP 2007.
Spectral surface reconstruction from noisy point clouds R. Kolluri (Google) J. Richard Shewchuk J. F. O’Brien University of Califonia, Berkeley SGP 2004
The eigencrust algorithm Partition the tetrahedra of a Delaunay tetrahedralization into inside/outside Identify the triangular faces that interface between the subgraphs
Poles Nina Amenta, Marshall Bern, Manolis Kamvysselis. A new Voronoi-based surface reconstruction algorithm. SigGraph 98
Pole graph G
The negatively weighted edges of the pole graph
Pole graph G The positively weighted edges of pole graph
Weights
Super node->G’
Pole matrix
Remaining tetrahedra
The final mesh The final mesh is the “eigencrust” The triangles where the inside and outside tetrahedra meet
Results If all adjacent tetrahedra are labeled the same, the point is an outlier Undersampled regions are handled without holes
More results
Efficacy input points Tetrahedralize:13.5 minutes 157 minutes 265minutes
Voronoi-based variational reconstruction of unoriented point sets P. Alliez D. Cohen-Steiner Y. Tong M. Desbrun SGP 2007 (best paper award)
Pierre Alliez Researcher at INRIA in the GEOMETRICA group Research Geometry Processing: geometry compression, surface approximation, mesh parameterization, surface remeshing and mesh generation Avid user of the CGAL library CGAL developer
David Cohen-Steiner Researcher at INRIA in the GEOMETRICA team Research Approximation problems in applied geometry and topology Meshes and point clouds are of particular interest
Yiying Tong Assistant Professor Computer Science and Engineering Dept. at Michigan State University Research Computer simulation/animation Discrete geometric modeling Discrete differential geometry Face recognition using 3D models
Mathieu Desbrun Associate Professor in Computer Science and Computational Science & Engineering California Institute of Technology Research Applying discrete differential geometry to a wide range of fields and applications
Overview Point set Tensor estimation Implicit function + contouring
Tensor estimation
Normal estimation(PCA)
Voronoi PCA
Noise-free case
Noise-free vs noisy
Noisy case
Implicit function Tensors
Delaunay refinement
Variational formulation Find implicit function f such that its gradient f best aligns to the principal component of the tensors Anisotropic Dirichlet energy Measures alignment with tensors f Biharmonic energy Measures smoothness of f Regularization
Rationale Anisotropic tensors: favor alignment Isotropic tensors: favor smoothness
Rationale Anisotropic tensors: favor alignment Isotropic tensors: favor smoothness Large aligned gradients + smoothness ->consistent orientation of f
Solver A: Anisotropic Laplacian operator B: Isotropic Bilaplacian operator Desbrun M, Kanso E, Tong Y. Discrete differential forms for Computational modeling. In Discrete Differential Geometry. ACM SIGGRAPH Course, V vertices { v i } E edges { e i } Tensor C F=(f 1,f 2,…,f v ) t
Solver
Generalized eigenvalue problem maxEigenvector (PWL function)
Standard eigenvalue problem Solver: Implicitly restarted Arnoldi method (ARPACK++)
Contouring F=(f 1,f 2,…,f v ) t
Sparse sampling
Noise
Nested components
Comparison PoissonGEP Poisson reconstruction
Comparison Poisson reconstruction
Sforz(250K points) Out-of-core factorization 25 minutes
Conclusion Pros Handles unoriented point sets Handles noisy point sets Cons Slow Not easy to implement