3d scattered data approximation with adaptive compactly...
TRANSCRIPT
![Page 1: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/1.jpg)
3D Scattered Data Approximation with Adaptive Compactly Supported RBFsDiego Salume
Authors: Ohtake, Belyaev, Seidel.
![Page 2: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/2.jpg)
Overview•Input: A set of points scattered over a piecewise smooth surface with oriented normals.
•Output: An implicit function whose zero level set is an approximation of the surface (robust to noise).
•Use compactly supported radial basis functions (RBFs) centered at randomly chosen points in the set.
•Support size for each RBF is adapted depending surface geometry around that RBF center.
•Takes into account confidence values for each sample point.
![Page 3: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/3.jpg)
RBFs•Given a set of N points pi with normals ni and confidence values vi .
•Want to find implicit function y=f(x), such that its zero level set approximates points pi .
•Given M approximation centers ci, such that M < N construct
Base approximation Local details
![Page 4: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/4.jpg)
RBFs•Our unknowns are gi(x) and λi.
◦ g is a local quadratic estimation of our points in (region of influence)
◦ Coefficients λi are determined from M interpolation conditions
◦ Base approximation term has the same zero level set as partition of unity approximations that use normalized RBFs, so we rewrite f
Adaptive PU Normalized RBF (refine PU approx)
![Page 5: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/5.jpg)
![Page 6: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/6.jpg)
Algorithm Outline
•Adaptive PU Approximation◦ Select approximation centers ci
◦ Assign influence parameters σi.
◦ Estimate function gi.
•Least Squares RBF Approximation◦Find coefficients λi
![Page 7: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/7.jpg)
Algorithm Outline
•Adaptive PU Approximation◦ Select approximation centers ci
◦ Assign influence parameters σi.
◦ Estimate function gi.
•Least Squares RBF Approximation◦Find coefficients λi
![Page 8: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/8.jpg)
Adaptive PU Approximation:Finding Local Quadratic Estimate g•To account for density irregularities, weight each point by the distance to its neighbors.
•Define local coordinate system (u,v,w) at center ci such that the positive w axis is the weighted average of the normals of the center’s σ-neighborhood
•Define local fitting function h
•Found by minimizing
![Page 9: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/9.jpg)
Algorithm Outline
•Adaptive PU Approximation◦ Select approximation centers ci
◦ Assign influence parameters σi.
◦ Estimate function gi.
•Least Squares RBF Approximation◦Find coefficients λi
![Page 10: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/10.jpg)
Adaptive PU Approximation:Finding Influence Parameters σ•To find optimal influence parameter σi associated with a center ci , we define error function
•Assume this function is monotonically decreasing to zero as σ approaches zero.
•Use Rissanen’s minimum description length (MDL) principle:
L: main diagonal of the bounding box of sample points1/L: for scale-independence
![Page 11: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/11.jpg)
Adaptive PU Approximation:Finding Influence Parameters σ•Find the approximation of a noisy signal that is the linear combination of the smallest number of approximants in a given collection.
•The distance from p to gi(x) = 0 is approx
•Elocal(σ)2 is proportional to the negative logarithm ⇔number of bits required to describe points near ci.
•Choose σ that minimizes
•C is a positive constant that controls the trade off between sparsity and approximation, and the smoothness of the reconstruction.
![Page 12: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/12.jpg)
Adaptive PU Approximation:Finding Influence Parameters σ•If σ is large, then the number of approximation centers is small and local error Elocal(σ) is large.◦ Thus, ESA(σ) grows drastically
as σ goes to infinity.
•For small σ, the number of approximation centers is large, since the zero level-set must reproduce noise.
![Page 13: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/13.jpg)
Adaptive PU Approximation:Finding Influence Parameters σ•The value of σi reflects the surface complexity at ci.◦ Bigger complexity ⇒ smaller σi.
•Two approaches for selecting σi:◦ Minimize ESA(σ): one-dimensional
problem. Penalize number of local approximations.
◦ Solve equation Elocal(σ) = 𝜖0 for a user-specified accuracy.• Small 𝜖0 ⇒ reconstruction of noise.
• Large 𝜖0 ⇒ oversmoothing.
![Page 14: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/14.jpg)
Algorithm Outline
•Adaptive PU Approximation◦ Select approximation centers ci
◦ Assign influence parameters σi.
◦ Estimate function gi.
•Least Squares RBF Approximation◦Find coefficients λi
![Page 15: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/15.jpg)
Adaptive PU Approximation:Selecting the centers ci•Choose centers so that their corresponding balls cover all the sample points with overlap greater than a threshold.
•The cover for center ci is a ball of radius σi centered at ci , defined by .
•Measure overlap at point pj as
![Page 16: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/16.jpg)
Adaptive PU Approximation:Selecting the centers ci•Algorithm for selecting ci with Toverlap:
1. Assign vj = 0 for each sample point pj.
2. Choose m random sample points with v < Toverlap.
3. Select the point with the minimum value of v.
4. Choose that point as an approx center ck and set vk = Toverlap.
5. Find support size σk and quadratic approximation gi.
6. Update overlap value vj for all sample points not selected as centers by adding
7. If there are points with v < Toverlap, go to step 2.
![Page 17: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/17.jpg)
Adaptive PU Approximation:Selecting the centers ci•Algorithm for selecting ci with Toverlap:
1. Assign vj = 0 for each sample point pj.
2. Choose m random sample points with v < Toverlap.
3. Select the point with the minimum value of v.
4. Choose that point as an approx center ck and set vk = Toverlap.
5. Find support size σk and quadratic approximation gi.
6. Update overlap value vj for all sample points not selected as centers by adding
7. If there are points with v < Toverlap, go to step 2.
![Page 18: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/18.jpg)
Adaptive PU Approximation:Noisy Data•Extra zero level-set surfaces appear on noisy datasets.
•To avoid artifacts, prevent influence parameter σ from being too small.◦ Modify Elocal(σ) for cases with small influence parameters:
![Page 19: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/19.jpg)
Algorithm Outline
•Adaptive PU Approximation◦ Select approximation centers ci
◦ Assign influence parameters σi.
◦ Estimate function gi.
•Least Squares RBF Approximation◦Find coefficients λi
![Page 20: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/20.jpg)
Least-Squares RBF Approximation:Determining RBF weights λi•Define global L2-error metric:
•Want to minimize global error metric to obtain RBF weights, but this produces overfitting.◦ Use a regularization approach to suppress oscillations. Modification:
◦ This is a quadratic min problem so
![Page 21: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/21.jpg)
Least-Squares RBF Approximation:Determining RBF weights λi•Define global L2-error metric:
•Want to minimize global error metric to obtain RBF weights, but this produces overfitting.◦ Use a regularization approach to suppress oscillations. Modification:
◦ This is a quadratic min problem so
![Page 22: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/22.jpg)
Discussion•Parameter selection
◦ TSA controls the smoothness of the model
◦ Toverlap and Treg are fixed.
◦ σmin = 0 for low noise.
◦ σmin = L/100 for noisy data.
![Page 23: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/23.jpg)
Discussion•Hole filling
![Page 24: 3D Scattered Data Approximation with Adaptive Compactly ...misha/Fall13b/Notes/Ohtake04.notes.pdf · Adaptive PU Approximation: Selecting the centers c i •Algorithm for selecting](https://reader035.vdocuments.us/reader035/viewer/2022070922/5fbac7dfec51234e7a7aecfe/html5/thumbnails/24.jpg)
Discussion•Performance
◦ To evaluate function f(x), we need to find all center ci such that xbelongs to their areas of influence.• Use a range searching octree-based data structure.
◦ For visualization, use Bloomenthal’s polygonizer.• One linear interpolation pass required to find f(x) = 0 since f(x) mimics the distance
function to the level set when close to zero.
◦ Time complexity and number of approximation centers depends on size of the dataset and on the geometric complexity.
•Future work◦ Selecting approximation centers that are not restricted to sample
points.◦ Combine method with a multi-scale approach.