Partial Shape Matching Outline: • Motivation • Sum of Squared Distances Representation Theory Motivation We have seen a large number of different shape descriptors: – – – – – – – – – – Shape Distributions Extended Gaussian Images Shape Histograms Gaussian EDT Wavelets Spherical Parameterizations Spherical Extent Functions Light Field Descriptors Shock Graphs Reeb Graphs Representation Theory Challenge Partial shape matching problem: Given a part of a model S and a whole model M determine if the part is a subset of the whole SM. S M Representation Theory Difficulty For whole object matching, we would associate a shape descriptor vM to every model M and would define the measure of similarity between models M and N as the distance between their descriptors: D(M,N)=||vM-vN|| Representation Theory Difficulty For partial object matching, we cannot use the same approach: – Vector differences are symmetric but subset matching is not: v M v N v N v M but S M M S – If SM, then we would like to have vS≠vM and: D(S , M ) 0 which means that we cannot use difference norms to measure similarity. Representation Theory Motivation We have seen a number of different ways for addressing the alignment problem: – – – – – Center of Mass Normalization Scale Normalization PCA Alignment Translation Invariance Rotation Invariance Representation Theory Motivation Most of these methods will change give very different descriptors if only part of the model is given. – Center of mass, variance, and principal axes of a part of the model will not be the same as those of the whole. Representation Theory Motivation Most of these methods will change give very different descriptors if only part of the model is given. – Changing the values of a function will change the (nonconstant) frequency distribution in non-trivial ways. Outline: • Motivation • Sum of Squared Distances Representation Theory Goal Design a new paradigm for shape matching that associates a simple structure to each shape M→vM such that if SM, then: – vS≠vM (unless S=M) – but D(S,M)=0 That is, we would like to define a measure of similarity that answers the question: “How close is S to being a subset of M?” Representation Theory Key Idea Instead of using the norm of the difference, use the dot product: D(S , M ) v~S ,v M v~ Then, S is a subset of M if S is orthogonal to v M . To do this, we have to define different descriptors for a model depending on whether it is the target or the query. Representation Theory Implementation For a model M, represent the model by two different 3D function: 1 if p M ~ v M RasterM ( p ) 0 otherwise 2 v M EDTM (p ) min p q q M M RasterM EDTM Representation Theory Implementation Then RasterM is non-zero only on the boundary points of the model, and EDTM is non-zero everywhere else. Consequently we have: RasterM (p ) EDTM (p ) 0 p and hence: v~M ,v M RasterM ( p ) EDTM ( p )dp 0. M RasterM EDTM Representation Theory Implementation Moreover, if SM, then we still have: RasterS ( p ) EDTM ( p ) 0 p so that: v~S ,v M RasterS ( p ) EDTM ( p )dp 0. S RasterS M EDTM Representation Theory What is the value of D(S,M)? Representation Theory What is the value of D(S,M)? D(S , M ) v~S ,v M RasterS ( p ) EDTM ( p )dp Representation Theory What is the value of D(S,M)? D(S , M ) v~S ,v M RasterS ( p ) EDTM ( p )dp Since RasterS is equal to 1 for points that lie on S and equal to 0 everywhere else: D(S , M ) EDTM ( p )dp p S Representation Theory What is the value of D(S,M)? D(S , M ) EDTM ( p )dp p S So that distance between S and M is equal to the sum of squared distances from points on S to the nearest point M. S M Representation Theory What is the value of D(S,M)? D(S , M ) EDTM ( p )dp p S Note that if we rasterize the models into an nxnxn voxel grid, then a brute force computation would compute the sum of the distances for each of O(n2) on the query by testing against each of O(n2) points on the target for the minimum distance, giving a total running time of O(n4). By pre-computing the EDT, we reduce the computation to O(n2) operations. Representation Theory Advantages Model similarity is defined in terms of the dotproduct: – We can still use SVD for efficiency and compression (since rotations do not change the dot product) – We can still use fast correlation methods (translation, rotation, axial flip) but now we want to find the transformation minimizing the correlation. Representation Theory Advantages We can use a symmetric version of this for whole Sum of Squared Distances (3D) 100% object matching. Spherical Extent Function (2D) Gaussian EDT (3D) Shape Histograms (3D) Extended Gaussian Image (2D) D2 (1D) 50% 0% Representation Theory Advantages We can perform importance matching by assigning a value larger than 1 to sub-regions of the rasterization. Representation Theory Disadvantage Aside from using fast Fourier / Spherical-Harmonic / Wigner-D transforms, we still have no good way to address the alignment problem.