520 likes | 780 Vues
Computational metric geometry. Michael Bronstein. Department of Computer Science Technion – Israel Institute of Technology. What is metric geometry?. ?. Metric space. Similarity of metric spaces. Metric representation. information retrieval. shape analysis. inverse problems.
E N D
Computational metric geometry Michael Bronstein Department of Computer Science Technion – Israel Institute of Technology
What is metric geometry? ? Metric space Similarity of metric spaces Metric representation
information retrieval shape analysis inverse problems medical imaging object detection Similarity
Non-rigid world from macro to nano Organs Proteins Nano- machines Micro- organisms Animals
Rock, paper, scissors Rock Scissors Paper
Rock, paper, scissors Hands Rock Scissors Paper
Invariant similarity Similarity Transformation
Metric model Shape Similarity Invariance metric space Distance between metric spaces and . isometry w.r.t.
‘ ‘ ‘ ‘ Isometry Two metric spaces and are isometric if there exists a bijective distance preserving map such that Two metric spaces and are -isometric if there exists a map which is • distance preserving • surjective -isometric -similar = In which metric?
Examples of metrics Euclidean Geodesic Diffusion
Rigid similarity Isometry between metric spaces Congruence Unknown correspondence! Min Hausdorff distance over Euclidean isometries
Non-rigid similarity Rigid similarity Non-rigid similarity Part of same metric space Different metric spaces SOLUTION: Find a representation of and in a common metric space
Canonical forms ? Compare canonical forms as rigid shapes Compute canonical forms Non-rigid shape similarity = Rigid similarity of canonical forms Elad, Kimmel 2003
Multidimensional scaling 7200 4000 5200 1630 TA SF 1800 Paris NY 2200 1900 2350 3100 Rio Find a configuration of points in the plane best representing distances between the cities
Multidimensional scaling Best possible embedding with minimum distortion Non-linear non-convex optimization problem in variables
Multigrid MDS Fine grid Solution Decimate Interpolate Relax Coarse grid Improved solution B et al. 2005
Multigrid MDS Execution time (sec) Multigrid MDS Standard MDS Stress Complexity (MFLOPs) B et al. 2005, 2006
Compute canonical forms (defined up to an isometry in ) Canonical forms, revisited Min distortion embedding Min distortion embedding Fix some metric space Compute Hausdorff distance between canonical forms No fixed (data-independent) embedding space will give distortion-less canonical forms!
Metric coupling Isometric embedding Isometric embedding Disjoint union ? ? How to choose the metric?
Gromov-Hausdorff distance Find the smallest possible metric Gromov-Hausdorff distance Distance between metric spaces (how isometric two spaces are) Generalization of the Hausdorff distance Gromov1981
Canonical forms Gromov-Hausdorff Fixed embedding space Optimal data-dependent embedding space Approximate metric (error dependent on the data) Metric on equivalence classes of isometric shapes -isometric -isometric Consistent to sampling for shapes sampled at radius
Gromov-Hausdorff distance Theorem: for compact spaces, is a correspondencesatisfying for every there exists s.t. for every there exists s.t. Optimization over all possible correspondences is NP-hard problem! Gromov1981
Multidimensional scaling Best possible embedding with minimum distortion
Generalized multidimensional scaling Best possible embedding with minimum distortion • Geodesic distances have no closed-form expression • No global representation for optimization variables • How to perform optimization on a manifold? B et al. 2006
GMDS: some technical details • No global system of coordinates Use local barycentric coordinates • No closed-form distances Interpolate distances from those pre-computed on the mesh • How to perform optimization? Perform path unfolding to go across triangles B et al. 2005
Canonical forms (MDS based on 500 points) Gromov-Hausdorff distance (GMDS based on 50 points) BBK, SIAM J. Sci. Comp 2006
Application to face recognition x y y’ x’ Euclidean metric
Application to face recognition x y y’ x’ Distance distortion distribution Geodesic metric
Eikonalvs heat equation Boundary conditions: Initial conditions: Viscosity solution: arrival time (geodesic distance from source) Solution : heat distribution at time t Kimmel & Sethian 1998 Weber, Devir, B2, Kimmel 2008
Heat equation on manifolds 1D 3D Heat kernel
Heat equation on manifolds 1D 3D Heat kernel “Convolution”
Diffusion distance “Connectivity rate” from to by paths of length • Small if there are many paths • Large if there are a few paths Geodesic = minimum-length path Diffusion distance = “average” length path (less sensitive to bottlenecks) Berard, Besson, Gallot, 1994; Coifman et al. PNAS 2005
Invariance: Euclidean metric Topology Rigid Scale Inelastic Wang, B, Paragios 2010
Invariance: geodesic metric Topology Rigid Scale Inelastic Wang, B, Paragios 2010
Invariance: diffusion metric Topology Rigid Scale Inelastic Wang, B, Paragios 2010
information retrieval shape analysis inverse problems medical imaging object detection Similarity
Metric learning “Similar” “Dissimilar” Generalization Data space Representation space Sampling of Metric learning: on training set
Similarity-sensitive hashing 0001 0011 0100 0111 1111 Data space Hamming space Shakhnarovich 2005 B2, Kimmel 2010; Strecha, B, Fua 2010
Video copy detection Lightsaber Luke vs Vader – Starwars classic Original copy Pirated copy
Mutation Biological DNA “Video DNA” So, what do you think? C C A A T T G C C C A AA T T G C C C C A A T TA G C C Substitution In/Del In/Del Substitution B2, Kimmel 2010
Mutation-invariant metric T So, what do you think? So, what do you think? So, what do you think? positive So, what do you think? So, what do you think? negative B2, Kimmel 2010
Video DNA alignment Gap Pairwise cost Optimal alignment = minimum-cost path Gap continuation Gap • Dynamic programming sequence alignment with gaps to account for In/Del mutations (Smith-WATerman algorithm) • Learned mutation-invariant pairwise matching cost B2, Kimmel 2010
Summary 0001 1001 1111 0111 1110 MDS Metric space Metric learning Object similarity is also a metric space Gromov-Hausdorff distance + GMDS Metric choice=invariance Examples of similarity (metric sampling)