A Diffusion Wavelet Approach For 3 D Model Matching
1. A Diffusion Wavelet Approach for 3-D Model Matching Authors: K.P. Zhu, Y.S. Wong, W.F. Lu, J.Y.H. Fuh Presented by: Raphael Steinberg
2. Schedule Introduction Diffusion Maps Wavelets and Diffusion Wavelets Fisher’s Discriminant Ratio (FDR) Retrieval Procedure Results Conclusions 2
3. Introduction Currently - A larger than ever number of 3D Models in CAD, computer games, multimedia, molecular biology, computer vision and more There is a need for 3D Retrieval 3
4. Introduction (2) Tagging are not always available or sufficient to describe the model we require Combine topological information with multi-scale properties 4
5. Model Reusability (CAD/Animation) Model Matching Video Retrieval (2.5D/Virtual environments) Ecommerce Correcting defects Efficient Representation Many other uses… Motivation for 3D Retrieval 5
6. Obstacles in Retrieval Partial retrieval - Non-transitive Functional description How to match text tags with vertices and texture? Orthonormal coordinate system 6
8. 3D Model Matching – Prior Art Feature vectors using wavelets to mesh vertices – localized in both space & frequency – Paquet et. al. 2000 Random sampling for comparison – Osada et. al. 2001 8
9. Spherical harmonics (SH) Global method in Euclidean space lacks multi-scale analysis Legendre polynomials solve the Laplace equation in Spherical coordinates Vranic et. al. 2001 9
10. Spherical Wavelets (SW) Multi-scale in Euclidean space Lacks connectivity on the manifold Tannenbaum et. al. 2007 10
11. Schedule Introduction Diffusion Maps Wavelets and Diffusion Wavelets Fisher’s Discriminant Ratio (FDR) Retrieval Procedure Results Conclusions 11
12. Diffusion Maps Introduction Originally suggested by Stephan Lafon and R.R. Coifman from Yale Math, circa 2005 Many other manifold learning techniques exist Data analysis based on geometric properties of the data set 12
20. Use RBF Gaussian Kernel to choose ε Normalize W to create a Stochastic Matrix Diffusion Maps Algorithm 16 Lu et. al. 2009
21. Diffusion Maps algorithm (2) Diffuse by taking higher powers of t “The diffusion distance is equal to the Euclidean distance in the diffusion map space” , Nadler et. al. 2005 Cut manifold according to dominant eigenvalues 17
22. Diffusion Maps Code Example function checker(); close all; tetha=2*pi*rand(1,500); z=[cos(tetha);sin(tetha)]; figure(1);scatter(z(1,:),z(2,:),'b*');hold on; N=size(z,2); epsilon=linspace(0.01,.3,10); %epsilon=.3; W=nan(N); summer=nan(1,length(epsilon)); for k=1:length(epsilon) for i=1:N parfor j=1:N W(i,j)=exp(-sum((z(:,j)-z(:,i)).^2)/2/epsilon(k)); end end summer(k)=sum(sum(W)); end figure;scatter(log(epsilon),log(summer));title('Epsilon - linear region') p=polyfit(log(epsilon),log(summer),1); d=2*p(1);%manifold dimension M=W*diag(1./sum(W,2)); [U V]=svds(M); sync=max(U(:,2)); figure(1);scatter(U(:,2)./sync,U(:,3)./sync,'rd') title('Original manifold as stars and reconstructed manifold as diamonds') end 18
23. Schedule Introduction Diffusion Maps Wavelets and Diffusion Wavelets Fisher’s Discriminant Ratio (FDR) Retrieval Procedure Results Conclusions 19
26. Novelty – Diffusion Wavelets Combination of Diffusion Maps and Wavelets Used for non-linear dimensionality reduction Extension of wavelets to the unit circle (just as diffusion maps extends the Fourier transform) 22
28. Example of Diffusion Wavelets 24 Wavelet basis ψ(2,2,3) Scaling basis φ(1,1,1) Wavelet basis ψ(4,2,5) Wavelet basis ψ(3,2,3)
29. Diffusion Wavelets Use an optimization scheme to construct the scaling functions Each scaling function should deal with a single dimension and be orthogonal to the other scaling functions Extension of wavelets to the sphere (or to any other manifold) 25
30. Diffusion Wavelets (2) Better than LOD (Level of Detail - simplifies meshes) Involved algorithm – very few implementations exist 26
37. IRPR Curve Measure performance – use Princeton University 3D database IRPR – Information Retrieval Precision-Recall 33
38. IRPR Curve m = relevant matches r = # of retrieved models 1) Precision = 2) Recall = 34
39. Schedule Introduction Diffusion Maps Wavelets and Diffusion Wavelets Fisher’s Discriminant Ratio (FDR) Retrieval Procedure Results Conclusions 35
40. 3D Model Retrieval Procedure Compute the diffusionwavelet for each 3D model Obtain the model representing vector X Compute the 2nd order statistics of X for each scale 36
41. 1) Start with a coarsest scale comparison 2)Advance up to the finest scale 3) Stop on threshold or when finest scale reached * Use a threshold to determine if a model is from a certain class Model Matching Procedure 37
42. Schedule Introduction Diffusion Maps Wavelets and Diffusion Wavelets Fisher’s Discriminant Ratio (FDR) Retrieval Procedure Results Conclusions 38
43. Experimental Results 39 Differences in scaling levels DW gives better results than SH and SW
45. Schedule Introduction Diffusion Maps Wavelets and Diffusion Wavelets Fisher’s Discriminant Ratio (FDR) Retrieval Procedure Results Conclusions 41
46. Authors’ Conclusions Surfaces with sharp peaks, grooves or holes contain high-frequency information which is not addressed by the wavelet multi-resolution (use diffusion wavelet packets instead?) Possible to extend to partial matching DW presents better results than SH and SW 42
52. Seems like a reasonable solution to the problem of 3D object retrieval44
53.
54. How are the wavelet functions affected when a new model is inserted?45
55.
56.
57. Can we have an extension of Diffusion Wavelets for non-rigid manifolds?47
58. “Would like to have” (Technical/2) How to automatically choose the level of decomposition 48
59. “Would like to have” (Technical/3) An intuitive explanation - why prefer Diffusion Wavelets over Diffusion Wavelet Packets? Wavelet Packets seem to give more information especially in high frequencies… 49
60. “Would like to have” (Technical/4) Numerical problems of overflow of the FDR - use logarithm instead of inverse? 50
61. “Would like to have” (Presentation/1) Block diagram of the algorithm 51
62. “Would like to have” (Presentation/2) Web-based Graphical User Interface 52
64. “Would like to have” (Presentation/4) More explanations on Diffusion Wavelets 54
65. Conclusions Shape retrieval requires multi-scale analysis 3D models, like most real-life objects, are embedded in a low dimension manifold Results are robust to noise and to mesh simplifications 55
66. Conclusions (2) Diffusion Wavelets give good retrieval results for 3D objects Possible to extend the proposed method to include texture, sound, smell, elasticity and any other possibly given attribute of the 3D model 56
68. References [1]K.P. Zhu, Y.S. Wong, W.F. Lu, J.Y.H. Fuh. , Department of Mechanical Engineering, National University of Singapore “A diffusion wavelet approach for 3-D model matching” Computer Aided Design, Elsevier, Nov. 2008 [2] Presentation by R.R. Coifman et. al. [3] J. Lu et. al. “Dominant Texture and Diffusion Distance Manifolds“, Eurographics, Volume 28 , Issue 2, Pages 667 - 676, Mar. 2009 [4] Diffusion waveletsMatlab code: http://www.math.duke.edu/~mauro/diffusionwavelets.html#Code|outline [5] The Princeton Shape Benchmark: http://shape.cs.princeton.edu/benchmark/ [6] Nadler, B., Lafon, S., Coifman, R., Kevrekidis, I. “Diffusion maps, spectral clustering and eigenfunctions of Fokker-Planck operators”. [7] Ulrike von Luxburg, “A tutorial on spectral clustering”. Statistical Journal 2007 [8] Personal communications with K.P. Zhu [9] MANI - Manifold learning Matlab tool http://www.math.umn.edu/~wittman/mani/ [10] Vranic D, Saupe D, Richter J. Tools for 3D-object retrieval: Karhunen-Loeve transform and spherical harmonics. In: Proc. IEEE workshop on multimedia signal processing; 2001. p. 29398. 58
69. References [11] Osada R, Funkhouser T, Chazelle B, Dobkin D. Matching 3D models with shape distributions, In: Proc. shape modeling international. 2001. p. 15466. [12] Laga H, Nakajima M. Statistical spherical wavelet moments for content-based 3D model Retrieval. In: Computer graphics international 2007, CGI. 2007; 2007. p.1-8. [13] Nain D, Haker S, Bobick A, Tannenbaum A. Multiscale 3-D shape representation and segmentation using spherical wavelets. IEEE Transactions on Medical Imaging 2007;26(4), pages 598-618. 59
Notes de l'éditeur
Google has their own format –SketchUp (SU) and has been investing a lot of effort in 3D technologies. More examples from google include – google earth, google sketch-up and O3D platform for browser 3D display
Other manifold learning techniques – PCA (+variants), Kernel PCA, LLE (locally linear embedding), ISOMAP, Hessian LLE, LaplacianEigenMaps and many more…
Reference[9]
Reference [2]
See [1,6] for diffusion maps equations and [3] for graph
Ignore λmax=1 since it is a trivial eigenvalue of the markov transition matrix and corresponds to the stationary distribution of the markov chain at t=∞Cut is promised to conform with min-cut max-flow algorithmSee reference [6] for a thorough description of diffusion maps
We see that it is possible to reconstruct the manifold using just a single eigenvalue. Cuts can be made on the eigenvectors that represent the manifold (in this case, the second or the third eigenvectors corresponding to the second or third largest eigenvalue) – these cuts are meaningful since they are taking into account the geometric distribution of the original points. We see that diffusion maps approximate the Fourier series over the circle as the sine and cosine functions are the solution of the differential equation f’’=-f
Animated objects can be more sensitive to mesh simplification algorithms than CAD models.
Haar wavelet with 3 levels of decomposition to the Stanford Bunny image. By applying a threshold in the wavelet domain we can efficiently find similar images. The threshold is a very efficient way to remove noise. The high value coefficients correspond to edges at various scales. Collecting the high value coefficients to create a feature vector would ensure a good representation of the image. For example, check the lossy compression algorithm JPEG-2000.
We take 4 levels of decomposition since there is no real advantage in taking more decomposition levels and the computational burden is heavy. This result is specific to Princeton’s database and can change when dealing with different databases.
Take inverse of FDR to avoid numeric problems of overflow. Select model with minimum within cluster scattering and maximum within cluster scattering
Can do training on the entire database…See [5] for database
Chair – sparse structurePlane – smooth surface with local singularityKangaroo – smooth surfaceFlower – combines smooth surface with local singularity