O PTIMAL S TEP N ONRIGID ICP A LGORITHMS

Brian Amberg, Sami Romdhani, and Thomas Vetter

{Brian.Amberg, Sami.Romdhani, Thomas.Vetter}@unibas.ch, University of Basel, Switzerland

P ROBLEM N ONRIGID -ICP: T RANSLATIONS R ESULTS : S YNTHETIC T ESTCASE

Reparametrize surfaces such that semantically corre- The simplest deformation parametrization is a displace- Map Template onto Target using N-ICP Affine, N-ICP Translation, Allen’03.
sponding points are mapped onto each other. ment vector per template vertex. We regularize by
minimizing the difference between the displacements of

Nonrigid
neighbouring vertices. This corresponds to minimizing
the gradient of the deformation field. In each step prelim-
inary correspondences are determined and the quadratic
function
    2
αM 0
E(X) = X − (2)
W In W (U − V ) F
is minimized.

Rigid
Map Template onto Scan giving a holefilled
U preliminary correspondence M node-arc incidence matrix of
Reparametrization. the template topology
V template vertices
X displacement vectors W diagonal weight matrix

M ETHOD N-ICP: A FFINE D EFORMATIONS

Global and local deformations of the mesh are recovered
by successive application of nonrigid ICP. The algorithm
start with a stiff template and successively relaxes the
stiffness to recover more local deformations.
Optimal steps are taken, when a unique deformation is
found for the chosen stiffness and correspondence.
This variant determines one 3 × 4 affine transformation
matrix Xi per vertex. The deformation is regularized by R ESULTS : FACE -S CANS
minimizing the difference between the transformations
acting on neighbouring vertices.
In each step preliminary correspondences are determined
and the quadratic function

Target
u3
    2
αM ⊗ G 0
u2
u1 u4 E(X) =  W D  X − W U  (1)
T

βD L UL F
S(X) X3 u5
X2 u6 u7 u8 is solved for the current stiffness α.
X4
v3
The three parts of the quadratic function correspond to
X1 v2
v4 X5 the stiffness, distance and landmark constraints.
X8

N-ICP Affine Result

v1 X6 X7 U preliminary correspondence X affine matrices
S v5 v8 D matrix build from template M node-arc incidence matrix of
v6 v7 vertices the template topology
U l , D l landmark matrices W diagonal weight matrix
The figure shows an intermediate step, with the template Template
S, the deformed positions S(X), the target T , and the
preliminary correspondences u.
ROBUSTNESS
E(X)
Residual

G ENERAL A LGORITHM
Initialize deformation parametrization X to the iden- 0 W ELL P OSEDNESS R EFERENCES
Iteration
tity transform. None of the submatrices M ⊗ G, D, and D L used for [1] B. Allen, B. Curless, and Z. Popović. The Space of Human Body
For each stiffness αi ∈ {α1 , . . . , αn }, αi > αi+1 To handle missing data the preliminary correspondences the Nonrigid ICP Affine method has full rank on its own. Shapes. TOG, 22(3):587–594, 2003.
are checked for validity using heuristics. The weights [2] B. Amberg, S. Romdhani, and T. Vetter. Optimal Step Nonrigid
j j−1 We prove in the paper that nonetheless the complete ma- ICP Algorithms for Surface Registration. In CVPR’07, 2007.
Until a fixpoint kX − X k < ε is reached W for invalid correspondences are set to zero, which trix has full rank and a unique solution for finding the [3] P. J. Besl and N. D. Mckay. A Method for Registration of 3-D
Find preliminary correspondences for deformed causes the residual to increase when more valid corre- optimal step exists. Shapes. PAMI, 14(2):239–256, February 1992.
j−1
template V(X ). spondences are found. Therefore a black-box optimizer Though this is only valid if the points are not coplanar [4] Y. Chen and G. Medioni. Object Modeling by Registration of Mul-
tiple Range Images. In CRA, volume 3, pages 2724–2729, 1991.
Determine X j as the optimal deformation for the is not usable in this setting. Convergence is determined our experiments showed that for faces the system is well [5] J. Feldmar and N. Ayache. Rigid, Affine and Locally Affine Reg-
preliminary correspondences and stiffness αi . by checking if a fix-point has been reached. conditioned. istration of Free-Form Surfaces. IJCV, 18(2):99–119, May 1996.