AmbergRegister: Register two triangular meshes based on smooth deformation.

In zarquon42b/mesheR: Meshing Operations on Triangular Meshes

Register two triangular meshes based on smooth deformation.

Description

Perform registration of two triangular meshes, minimizing per-face distortions.

Usage

AmbergRegister(
  x,
  mesh2,
  lm1 = NULL,
  lm2 = NULL,
  k = 1,
  lambda = 1,
  iterations = 15,
  rho = pi/2,
  dist = 2,
  border = FALSE,
  smooth = TRUE,
  smoothit = 1,
  smoothtype = "t",
  tol = 1e-10,
  useiter = TRUE,
  minclost = 50,
  distinc = 1,
  rigid = NULL,
  similarity = NULL,
  affine = NULL,
  tps = FALSE,
  pcAlign = FALSE,
  nn = 20,
  silent = FALSE,
  useConstrained = TRUE,
  forceLM = FALSE,
  visualize = FALSE,
  folder = NULL,
  noinc = FALSE,
  angclost = FALSE,
  bboxCrop = NULL,
  threads = 0
)

Arguments

x	reference mesh: triangular mesh of class "mesh3d"or of class BayesDeform created by createBayes to restrict based on a known distribution. To use this option the package RvtkStatismo https://github.com/zarquon42b/RvtkStatismo has to be installed.No loose vertices, edges and degenerated faces are allowed.
mesh2	target mesh: triangular mesh of class "mesh3d".
lm1	m x 3 matrix containing correspondences on "mesh1".
lm2	m x 3 matrix containing target correspondences on "mesh2".
k	integer: parameter regularizing face normal distortion. Can be vector of length(iterations) or single value.
lambda	numeric: parameter regularizing faces's distortion. Can be vector of length(iterations) or single value.
iterations	integer: number of iterations to run.
rho	numeric: 0 < rho < 2*pi tolerance of normal deviation between reference vertices and corresponding closest points on target suface.
dist	numeric: tolerance of maximal distance between reference vertices and corresponding closest points on target suface.
border	logical: if FALSE, hits on border faces are ignored (reduces distortion)
smooth	logical: if TRUE after each iteration a mesh smoothing is performed.
smoothit	integer: determine smoothing iterations.
smoothtype	character: select smoothing algorithm - see vcgSmooth for further details.
tol	numeric: convergence threshold of MSE between vertices of two successive iterations.
useiter	logical: if TRUE, each iteration uses the updated reference mesh, if false. The original mesh will be deformed based on the updated correspondences.
minclost	minimum amount of correspondence points. If less correspondences are found, dist will be increased by "distinc" (see below).
distinc	increment of dist, in case minclost is not reached.
rigid	named list. Passing parameters to icp, for rigid registration. If landmarks are provided and only those should count, set rigid$iterations=0.
similarity	named list. Passing parameters to icp, for similarity registration (rigid +scaling). If landmarks are provided and only those should count, set similarity$iterations=0 (and rigid=NULL).
affine	named list. Passing parameters to icp, for affine registration. If landmarks are provided and only those should count, set similarity$iterations=0 (with rigid=NULL and similarity=NULL)
tps	logical: if TRUE and landmarks are provided, the reference will be mapped to the target using a Thin-Plate Spline interpolation. Overrides rigid,affine and similarity.
pcAlign	if TRUE, surfaces are prealigned by principal axis. Overrides intial landmark based alignment.
nn	integer: closest barycenters. During search for closest points on target, the closest nn faces are probed. The larger nn is , the more accurate the closest point search but also the more time consuming. If landmarks are provided and only those should count, set rigid$iterations=0.
silent	logical: no verbosity
useConstrained	logical: if TRUE and Bayes and landmarks are defined, the landmarks are not only used to get a suitable reference but the model will also be constrained by the landmarks to subsequently restrict the shape variability. If FALSE, the full model is used.
forceLM	logical: if icp is requested landmark based deformation will be applied after icp-based transformation.
visualize	logical request visualization of deformation process.
folder	logical: if visualize=TRUE, this can specify a folder to save screenshots of each deformation state, in order to create a movie or an animated gif.
noinc	logical: if TRUE and x is of class 'Bayes', the process stops if the distance from the target to the deformed reference increases compared to the previous iteration.
angclost	if TRUE, the closest k faces will be evaluated and the closest with the appropriate normal angle will be selected.
bboxCrop	extend of the bounding box around mesh1 (after alignmend) that will be cropped from target to speed things up.
threads	integer: threads to use in closest point search.

Details

This function runs an elastic-ICP surface matching algorithm, that minimizes the original meshes internal structure by solving a sparse equation system. The user can control 2 parameters of mesh stiffness: lambda and k. lambda controls the impact of the control points (closest points) as it is a weight applied to the equation system. The value of lambda should be carefully selected depending on the object overall size: i.e. to match two tiny meshes one will need a higher value than a for a larger object (example: I found values between 0 and 1 suitable for human faces and values between 10 and 100 suitable for mice teeth). k controls the normal slackness, i.e. the deviation of normal direction. The larger, k, the more elastic the deformation will be. lambda and k can be specified as vectors of length iterations, to assign a specific value for each iteration.

Value

mesh	registered mesh
affine	affine 4x4 transformation matrix mapping mesh1 onto mesh2
lm1	lm1 mapped onto the registered template

Author(s)

Stefan Schlager

References

Amberg, B. 2011. Editing faces in videos, University of Basel.

See Also

gaussMatch

Examples

require(Morpho)

require(Rvcg)
data(humface)
data(dummyhead)
## set parameters making each iteration more elastic
# only 10 iterations to keep example calculation time reasonable.
params <- list(iterations=10) 
params <- append(params, list(
   # first lambda is set relatively high because first matching uses landmarks
   # then let it increase from 0.2 to 0.6
   lambda=c(0.7,seq(from = 0.2,to=0.6,length.out = params$iterations-1)),
   # treat k similar as lambda
   k=c(10,seq(from = 1,to=params$iterations-1,by=1)),
   useiter=FALSE # iteratively deform dummyhead onto humface
   ))
#we also want the landmarks to be used in an initial similarity transform
similarity <- list(iterations=0)
map <- AmbergRegister(dummyhead.mesh, humface, lm1=dummyhead.lm,
                 lm2=humface.lm, iterations=params$iterations,similarity=similarity,
                 k=params$k, lambda=params$lambda, useiter=params$useiter)
# compare matched and original face:
## Not run: 
require(rgl)
meshDist(map$mesh, humface ,from=-3,to=3,tol=0.5)
# render original mesh as wireframe
shade3d(humface,front="lines",back="lines")

## End(Not run)
##example with different icp matchings:
rigid <- list(iterations=30,subsample=200,rhotol=pi/2,uprange=0.6)
similarity <- list(iterations=30, subsample=200,rhotol=pi/2,uprange=0.6)
affine <- list(iterations=30,subsample=200,rhotol=pi/2,uprange=0.6)
map <- AmbergRegister(dummyhead.mesh, humface, lm1=dummyhead.lm,
                      lm2=humface.lm, iterations=params$iterations,
                      k=params$k, lambda=params$lambda, useiter=params$useiter,rigid=rigid,
                      similarity=similarity,affine=affine,forceLM = TRUE)

zarquon42b/mesheR documentation built on Jan. 28, 2024, 2:17 p.m.