Adaptive Multigrid

This page documents developments on the design and implementation of an adaptive multigrid solver that allows for the efficient solution of Poisson-like systems of equations, discretized over an quad/oct-tree.
This material is based upon work supported by the National Science Foundation under Grant No. 1422325:

Goals and Challenges

In numerous fields, ranging from image processing to scientific simulation, solving a large Poisson system is an essential step in integrating locally defined properties into a consistent global solution. Though the linear system is global (with properties at one point in space affecting the solution at points far away), in many cases the details of the solution are only required at particular locations. As a result, computational effort expended on estimating the details of the solution away from these regions of interest is wasted. The proposed research seeks to design, implement, and make publicly available a new solver for addressing this problem - providing a way to adapt traditional hierarchical solvers so that computation is only focused where it is needed.

The motivation for this approach stems from the fact that in solving a Poisson equation, long-distance effects are dominated by lower frequencies, so high-resolution components of the solution need only be computed in regions of interest. Using experiences from prior work in surface reconstruction and large image processing, the PI will explore a modification of traditional multigrid, providing a hierarchical solver even when the function spaces do not nest. To do this the multigrid solver will be modified in two ways. First, to compensate for the fact that the coarser solution cannot be up-sampled to the finer resolutions, the solution of the linear system will be represented as a linear combination of functions at all levels of the hierarchy, not just the finest one. Second, instead of using traditional restriction and prolongation operators to transition between successive levels of the hierarchy, the solver will proceed by iterating through the levels of the hierarchy (fine-to-coarse and then coarse-to-fine as in a typical V- or W-cycle), relaxing the system within each level, and adjusting the constraints at the other levels, accounting for the components of the solution that have already been met. (For example, in the prolongation phase, the coarser solution will be incorporated into the finer levels by adjusting the right-hand-side, rather than updating the current estimate of the solution.) The research will explore both the general design of such a system, as well as domain-specific implementations that facilitate its integration in targeted scientific and entertainment disciplines. Results will be made publicly available in the form C/C++ source code for constructing adapted quadtrees/octrees, formulating the system constraints, and evaluating the computed solution, as well as documentation and application-specific implementations.

Publications, Code, and Presentations

Direct Results

Initial research will focus on extending the cascadic solver used in Poisson Surface Reconstruction. This includes:

Indirect Results

Accurate Isosurface Interpolation with Hermite Data
This work considers the use of non-linear interpolants for iso-surface extraction (applicable to both reguar grids and adapted octrees) by considering not only the values at the end-points of an edge but also the gradients. This has been shown to be particular valuable for extracing high-quality surfaces when the underlying implicit function fails to be linear near the iso-surface (e.g. for indicator funcionts).
Gradient Domain for Color Correction in Large EM Stacks
This work considers the problem of performing gradient domain color-correction on huge (teravoxel) 3D EM data. The processing is decomposed into two phases: An initial smoothing pass to obtain a 3D dataset that is coherent along the z-axis, followed by independent 2D gradient domain fusion to obtain slices that preserve the high-frequency content of the input slices while exhibiting the coherence of the smoothed data. Separating the color correction into these two phases makes the approach trivially parallelizable and provides the space- and time-efficiency required for processing large 3D volumes.
Fast and Exact (Poisson) Solvers on Symmetric Geometries
This work considers the problem of solving linear systems on geometries with symmetries. In particular, when the linear system commutes with the symmetry group, the decomposition of the function space into irreducible representations results in a block-diagonalization of the matrix, replacing the solution of one large system with the solutions of many small ones. For Poisson equations on surfaces of revolution, the decomposition into irreducibles is obtained by computing a set of FFTs along the parallels of the surface, and the diagonal blcoks are tri-diagonally banded so that they can be solved in linear time.
Motion Graphs for Unstructured Textured Meshes
This work explores extensions of the Poisson equation to applications in vector-field design. Specifically, using the Laplacian of vector-fields as a regularizer, we design a hierarchial optical-flow solver for signals defined on surfaces that do not (directly) support a multi-resolution representation in order to obtain a good (non-local) minimizer to the non-convex and underconstrained optical flow probleem.
Field-Aligned Online Surface Reconstruction
This work explores the localization of updates within a hierarchical solver. Applied to the context of surface reconstruction, this allows a user to incrementally add new scans to a 3D reconstruction in time that is proportional to the size of the newly added scan, and independent of the overall size of the model. This, in turn, enables the design of an on-the-fly surface reconstruction system that interactively displays the reconstruction from the current scans and guides the user in choosing locations/orientations for subsequent scans.
Spatiotemporal Atlas Parameterization for Evolving Meshes
This work continues the work on "Motion Graphs", leveraging the optical flow solver to refine registration between two surfaces by using the optical flow of textures to constrain the tangential motion of the surface deformation.

Contact Info
Michael Kazhdan
Department of Computer Science
Johns Hopkins University
229 Malone Hall
3400 North Charles St.
Baltimore, MD 21218
Fax: (410) 516-6134

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation
Last updated 07/06/2017