List of required and optional parameters

Deformation and optimization parameter settings (in paramDiffeos.xml type of files)

The only compulsory parameter is kernel-width.

  • kernel-width: the width of the deformation kernel. The larger the more “rigid” the deformation. The smaller, the more local variations of the space is allowed.
  • kernel-type: exact (default), p3m, fgt. The method for evaluating kernels.
  • sparsity-prior (default 0.0): weight of the L^1 type prior.
  • number-of-threads (default 1): number of threads used. Only sparseAltas is multi-threaded. Each subject may be processed in a different thread.
  • smoothing-kernel-width-ratio (default 1.0): template parameters in sparseAltas are updated via a Sobolev gradient, which is computed thanks to a kernel whose width equals smoothing-kernel-width-ratio times kernel-width.
  • number-of-timepoints (default 10). Number of time steps in which the unit interval is discretized. It used for the integration of the differential equations.
  • initial-cp-spacing (equals kernel-width by default): the spacing of the regular lattice of control points, which is used as initialization of control points position in sparseMatching and sparseAtlas.
  • freeze-cp: On, Off (default). Optimizes or not the position of control points during optimization.
  • initial-momenta (default void): name of a file containing values of momentum vectors, which will be used as initial conditions in the optimization. If set, a file name in initial-cp-position should be given, otherwise this file is ignored.
  • initial-cp-position (default void): name of a file containing positions of control points, which will be used as initial conditions in the optimization. If no file name is given in initial-momenta, momentum vectors are initialized to zero. In case of conflict with initial-cp-spacing. If a file name is given in initial-cp-position and initial-cp-spacing is set, the latter is ignored and control point positions in the file name are used.
  • optimization-method-type: GradientDescent, ISTA, FISTA (default). If sparsity-prior is different from 0, GradientDescent is automatically switched to ISTA. If sparsity-prior = 0, FISTA is the Nesterov scheme, and ISTA is the same as GradientDescent.
  • max-iterations (default 100): maximum number of iterations during the optimization. To be increased if the convergence is slow.
  • adaptive-tolerance (default 1e-4): optimization stops if (E_curr – E_init) < adaptive-tolerance * E_init, where E_curr and E_init denote the current and initial value of the criterion respectively, and if the maximum number of iterations is not reached.
  • max-line-search-iterations (default 10): maximum number of iterations for adapting the stepsize during one step of the optimization method, using a line search method.
  • step-expand (default 2.0): factor by which the stepsize is multiplied after a success in the line search method.
  • step-shrink (default 0.5): factor by which the stepsize is multiplied after a failure in the line search method.
  • initial-step-multiplier (default 10): factor by which the initial stepsize is multiplied. The initial stepsize is determined by heuristics.
  • p3m-padding-factor (default 3): defines a padding of the regular lattice used in p3m optimization. The lattice is extended by p3m-padding-factor times kernel-width, to avoid side effects due to the periodic boundary conditions of the FFTs.
  • p3m-working-spacing-ratio (default 0.2): the ratio between the lattice step used in p3m optimization and the kernel-width. 0.2 gives an approximation error of about 5%, set to 0.3 for increased speed and an approximation error of about 10%.

Object parameter settings (in paramObject.xml type of files)

The first two parameters are compulsory, and the third one for objects of type other than Landmark.

  • deformable-object-type: Landmark, PointCloud, OrientedPolyLine, NonOrientedPolyLine, OrientedSurfaceMesh, NonOrientedSurfaceMesh. The type of object, which also determines the metric used measure similarity between objects.
  • data-sigma: the weight \sigma_k of this object in the cost function in sparseMatching and sparseAtlas
  • kernel-width: kernel-width for the metric used for objects of type PointCloud, OrientedPolyLine, NonOrientedPolyLine, OrientedSurfaceMesh and NonOrientedSurfaceMesh, namely objects using the currents or varifold representation.
  • reorient-normal: On, Off (default) uses (or not) a VTK filter to re-orient normals consistently for genus zero surface mesh. Use with care for surface mesh with holes or with several connected components. This flag is ignored for objects other than OrientedSurfaceMesh.

 

Comments are closed.