Shape reconstruction and inverse problems Lecture 9 © Alexander & Michael Bronstein tosca.cs.technion.ac.il/book Numerical geometry of non-rigid shapes Stanford University, Winter 2009 1
Measurement Projection Shape space Measurement space
Reconstruction ? Find Given Shape space Measurement space 4
Reconstruction Shape space Measurement space 5
Given a (possibly noisy) measurement of unknown shape Inverse problems Given a (possibly noisy) measurement of unknown shape Reconstruct the shape by minimizing the distance between given measurement and measurement obtained from shape 6
Inverse problems 7
Many shapes have the same measurement! Ill-posedness Many shapes have the same measurement! Shape space Measurement space 8
Prior knowledge We know that the measurements come from deformations of the same object! 9
Regularization Deformations of the dog shape Shape space Measurement space 10
Regularization Prior Shape space Measurement space 11 Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 11
Inverse problems with intrinsic prior Error Regularization Prior is given on the intrinsic geometry of the shape (intrinsic prior) Error = distance between measurements Regularization = intrinsic distance from prior shape Prior shape is a deformation of the shape we need to reconstruct Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 12
Solution of inverse problems with intrinsic prior Optimization variable: shape , represented as a set of coordinates Possible initialization: prior shape Gradients of and w.r.t. are required Does it sound familiar? Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 13
? Extrinsic dissimilarity Intrinsic dissimilarity
Joint similarity as inverse problem Measurement space = shape space Identity projection operator = intrinsic distance on shape space = extrinsic distance on measurement space Prior = the shape itself 15
Computation of the regularization term Assume shape = deformation of prior with the same connectivity Trivial correspondence Compute L2 distortion of geodesic distances and gradient is a fixed (precomputed) matrix of geodesic distances on depends on the variables (must be updated on every iteration) A. Bronstein, M. Bronstein, R. Kimmel, ICCV 2007
Computation of using Dijkstra’s algorithm Same approach as in joint similarity Compute and fix the path of the geodesic is a matrix of Euclidean distances between adjacent vertices is a linear operator integrating the path length along fixed path At each iteration, only changes Computation of is straightforward A. Bronstein, M. Bronstein, R. Kimmel, ICCV 2007
Inconsistency of Dijkstra’s algorithm Dijkstra/Analytic 1.1 1.05 FMM/Analytic 1 Number of points Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009
Computation of using Fast Marching Standard FMM FMM with derivative propagation Distance update Distance update Distance derivative update Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009
Shape-from-X Silhouette Sparse points Image Shading Stereo 20
Denoising Measurement space = shape space Identity projection operator = intrinsic distance on shape space = extrinsic distance on measurement space Noisy measurement 21
Denoising Unknown shape Prior Measurement Reconstruction (without prior) Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 22
Denoising Unknown shape Prior Measurement Reconstruction (with prior) Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 23
Bundle adjustment Clean measurement Shape Noisy measurement 24 Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 24
Bundle adjustment Measurement space of 2D point clouds Projection operator (assuming known correspondence) Noisy measurement 25
Bundle adjustment Unknown shape Prior Measurement Reconstruction (without prior) Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 26
Bundle adjustment Unknown shape Prior Measurement Reconstruction (with prior) Devir, Rosman, Bronstein, Bronstein, Kimmel, 2009 27
Shape-to-image matching Measured image Reconstruction Salzmann, Pilet, Ilic, Fua, 2007 28