978-1-4673-9098-9/15/$31.00 ©2015 IEEE 504
2015 8th International Congress on Image and Signal Processing (CISP 2015)
An Algorithm for Dense Correspondence Based on
Blended Intrinsic Map
Dan Kang
1,2
, Xiuyang Zhao
1,2
, Zhiang Chen
1
, Mingjun Liu
1,2,
∗
1
School of Information Science and Engineering, University of Jinan, Jinan 250022, PR China
2
Shandong Provincial Key Laboratory of Network Based Intelligent Computing, Jinan 250022, PR China
Abstract—The dense matching of 3-D meshes is an important
research topic in the field of computer vision. In this paper,
we present a layered matching pipeline based on the mixed
corresponding grid dense matching algorithm. Firstly, this al-
gorithm find an intrinsic map between two non-isometric, genus
zero surfaces. Secondly, we use the nature of the bottom of the
measure preserving distance to released the dense corresponding
of surfaces. A set of experimental results show that the proposed
method achieves better approximation accuracy than the ICP
method.
Index Terms—Mesh Dense correspondence; Layered matching;
1-ring points;
I. INTRODUCTION
With the development of technology, the three-dimensional
digital geometry draw widespread attention, at the same time,
it has been widely used in a wide range of industries
[1]
,
such as in the culture relic restoration, 3-D entertainment,
industrial design and so on. Therefore, the research of the
three-dimensional digital geometry has more and more pro-
found meaning and becomes a hot research topic in computer
graphics
[2]
.
Shape corresponding is the basis of 3-D digital geometry
processing technology
[3]
, and has widely used in the areas
such as model retrieval
[4]
, deformation, file and television
production, mesh game and so on. The question of shape
corresponding can described as follows: Given a set of ge-
ometric models such as S1,S2, ... ,SN, our aim is to look for a
meaningful mapping or correspondence between the geometric
models
[5]
.
Finding correspondences between surfaces problem has a
rich variety of literature. Previous approaches can be sorted by
the method used to find correspondences
[6]
. In any case, the
aim is to find a smooth, low-distortion (possibly partial) map
between significantly non-isometric surfaces in polynomial
time. One of the most classical method of search for corre-
spondences is the method of Iterative Closest Point (ICP). The
core idea of ICP is that iterative to search for the closest point
on the target mesh which has least Euclidean distance with
a point on source mesh
[7,8]
. However, when the differences
of position and orientation are large between the target mesh
and the source mesh, the ICP algorithm may converge to a
local extreme point
[9]
. The algorithm is mainly using the local
shape similarity model points corresponding to the problem.
∗
Corresponding author.lmj@ujn.edu.cn
Compared with the shape of the global information, the
method of local similarity is in poor stability. Therefore, when
use method of ICP, the initial meshes should be in similarity
position and orientation.
Multidimensional Scaling(MDS)
[10]
. For the approxima-
tion of non rigid correspondence problem between isometric
meshes, transform a model from Euclidean space into a m-
dimension space by space transformation
[11]
, while the mesh-
es have invariants geometric in the new space, eventually
transform problem from nonrigid correspondence into rigid
correspondence which is in the new space
[12]
. Making use
of the Euclidean distance in m-dimension approximate the
Euclidean distance in three dimension
[13]
. Method of MDS
makes source mesh and approximate isometric model mesh
embed into a lower dimension space
[14]
, and then transforms
problem into a rigid correspondence
[15]
. The method almost
is used in a relatively simple model and always has effective
results. However, this kind of method is not only time costum-
ing, but also is sensitive to initial, and unstable for the mesh
that has rich characteristics.
In recent years, the thermal diffusion method is widely
used in 3D shape recognition. Heat kernel signature (HKS)
was introduced in 2009 by Jian Sun, et al
[16]
. It is based
on heat kernel, which is a fundamental solution to the heat
equation
[17]
. HKS is one of the many recently introduced
shape descriptors which are based on the Laplace-Beltrami
operator associated with the shape. HKS is also a feature
descriptor for use in deformable shape analysis and belongs to
the group of spectral shape analysis methods
[18,19]
. For each
point in the shape, HKS defines its feature vector representing
the point’s local and global geometric properties. The appli-
cation areas of the HKS include segmentation, classification,
structure discovery, shape matching and shape retrieval. This
kind of algorithm mostly use local similarity shape to solve
point correspondence, but compared with the global shape
information, it has poor stability
[20]
.
Because there are rare single methods which can realize a
mesh corresponds correctly in every point in meshes with low
distortion, the general approach to this problem is to looking
for more groups of good maps, and then blend all the maps
with some weights to realize low-distortion everywhere
[16]
.
However, because the method has many limitations and for
reasons of computational efficiency, the number of mappings