Computer Graphics Project Report: Usama Mehmood - 110614650 Stony Brook University

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Computer Graphics Project Report

Usama Mehmood - 110614650


Stony Brook University

Abstract correspondences. [3], [2], [7] and [4] assume


a rigid transform between the models, whereas
The problem of shape matching and surface [9], [8], [5] and [1] also cater to non-rigid
registration has gained a lot of attention lately. deformations such as isometries.
Various algorithms have been proposed for A deformation is isometric if it transforms the
this purpose. Most techniques sub-sample the surface without changing its intrinsic proper-
points on the shape for efficiency. Sampling ties. One such property is geodesic distances.
technique that captures maximal information So an isometry does not stretch or tear the
content has an edge. Geodesic distances be- surface. Many deformations in application
tween sampled points are used to create shape can be approximated well by isometries. For
descriptors. Rigid body Registration problem example moving-parts of machines, human
is solved to a great degree of accuracy. It is poses and facial expressions.
matching under deformations that is currently For most applications the 3d models are de-
being tackled enthusiastically. The paper I im- tailed and hence they have a huge vertex count.
plemented in this project gives a fast algorithm Most of the low level detail can be ignored
for matching under isometric deformations. for the matching problem. Hence for the sake
of efficiency, the first step of matching is to
sub-sample the surface. [9] uses Poisson disk
1 Introduction sampling, [5] uses uniform sampling and [8]
uses farthest point sampling algorithms. The
Due to a surge in modelling and acquisition
sampling scheme ensures that the distribution
techniques, 3d models are becoming common
of points captures the shape of the object
in many fields such as entertainment, computer
without a significant loss of information. For
aided design, architecture and medicine.
rigid transforms it makes more sense to choose
Hence, the problem of shape matching and
a small subset of points because rigid transform
surface registration has gained a lot of attention
can be specified using very few parameters.
lately.
[9], [8], [5] and [1] make use of geodesic
Given two 3d models that are related by a
distances for the matching. Geodesic dis-
transformation the goal of shape matching is
tance is invariant to isometric deformation.
to get a dense correspondence between points
Euclidean distance although works well for
on both the shapes. Surface registration on
rigid body transforms, is not conserved when
the other hand tries to find the transformation
the bodies undergo a non-rigid transform. All
between the models ,in addiction to point
these methods incorporate geodesic distances
in the computation of feature descriptor. Each
method has its own pros and cons as discussed
in the following sections. But first we will
look at the Rigid body registration problem.

Isometrically deformed shapes. The geodesic (a) Model number one, (b) Model number two
distances as well as geodesic paths do not
change under such a deformation

3 Poisson disk sampling

First step in the matching pipeline is to


sub-sample both the input models to reduce
the number of points. Multiple techniques
2 Project Overview [4] like farthest point sampling and poisson disk
sampling could be used. I implemented a
Poisson disk sampler. It enforces that no two
points on the resampled shapes are closer than
This is a sampling based shape matching a predefined ε. The value of ε dictates the
algorithm that has a planning step to find number of points that are ultimately sampled.
”landmark” points. These points contain high- I use a value of five percent of the diameter
est information content about the shapes. This of input model as my value for ε. Using this
planning set is matched first, and then dense value I get 216 points on Model 1 and 194 on
correspondences are computed. Algorithm model 2. It is important to note that both the
works on meshes as it utilizes the connectivity models had 2503 points originally.
information. This technique does not depend The poisson disk sampler randomly chooses
upon key points. They argue that key-point a point and removes all other points that are
based techniques are a severe limitation as at the most ε distance away from the model.
they fail on models with insufficient surface All the distances are geodesic. They are
features. Hence this algorithm has an edge over computed on the mesh using an open source
the techniques discussed above. I have carried implementation of djikstra’s algorithm [6].
out my experimentation on the famous stanford It continues iteratively until all the points are
bunny. I try to match points points after the either included or removed. Sampled points
model undergoes rigid body transform. Input are shown in the following figures.
models are represented in the following figures.

2
about the shape. They are sparsely distributed.

4 Iso-curve calculation
Isocurves defined around a point on the mesh
are used to define local descriptors used in this
project. An isocurve around point x is a set of
Samples on Model number one points that are a fixed distance away from x.
I compute an isocurve by finding the geodesic
distance from the source x to all target points.
Then I include only the points that fall fixed
distance away from the source. These are the
points that lie on the isocurve. The only param-
eter used to define an isocurve around a partic-
ular point is the distance ρ.
Given a Set C which contains points that are on
an isocurve at ρ from x, I compute the length of
isocurve by iteratively finding the distance be-
tween all adjacent points in C. For a point m,
adjacent point is n ε C that is closest to m.
Samples on Model number two

For the purpose of experimentation I tested


with ε equal to 2.5 percent of the diameter
of the model. It samples only 61 points on Isocurves on the surface.
the surface. Rusults are presented below for
comparison.
5 Descriptors
A descriptor for a particular point is a set of
normalized lengths of isocurves that lie on the
mesh. Length is measured within surface of the
mesh. This is referred as geodesic distances.
An isocurve for a given point is as a set of
points that are fixed distance apart from the
given point. For descriptors they I chose 16
isocurves in the range [0 ρmax ]. ρmax is the
only parameter that can be adjusted in this
scheme. A smaller value focuses on local
Samples on Model with ε = 0.025d information whereas larger value includes
global properties. ρmax is set to thirteen percent
As it can be seen ,with this value of ε ,the of the diameter of model.
sampled points cover inadequate information Descriptors are calculated for all 216 points for

3
model 1 and 193 points for model 2. They are
stored in a file for further use. The computation
of descriptors is a computationally expensive
task. It took four hours to compute descriptors
for four hundred points.

Position of Vertex number 1221 and 1926.

6 Matching Algorithm
Comparison of descriptors on the tip of bunny’s
ear. (a) Vertex number 1926, (b) Vertex Num- The following probbabilty values are needed
ber 950, diff is the difference between both for the matching algorithm. (1) is the prob-
descriptors. abilty of how good the descriptors Dx and
Dy match. It has a parameter σd which is
set to 0.55 and σg used in (2) which is set
In the above figure both the descriptors are to 0.045. These values dictate the sensitivity
for points are the tip of the bunny’s ear. Hence of the probability values to the difference in
the error between them is really low. Following descriptor values. If these parameter values are
figure displays contrasting descriptors. increased then the probabilities get closer to
zero. There is a room for experimentation to
find the optimum values.

Pdist on the other hand uses the previously


matched point correspondences to find the
probabilty of how good current points match.

Psamp is a product of Pdist and Pdesc . It


combines both probabilities in a product. Psamp
is used finally to see how good a point y in
Comparison of descriptors on the tip of bunny’s shape 2 matches to a point x in shape 1.
ear and tip of feet. (a) Vertex number 1221,
(b) Vertex Number 1926, diff is the difference Pdist requires the computation of geodesic
between both descriptors. distances. If computed on runtime the algo-
rithm takes too long too compute. This is be-
cause geodesic distances on the mesh are com-
The position on the bunny for the vertices puted using the Dijkstra’s algorithm. Dijkstra’s
whose descriptors are shown in the above fig- algorithm is computationally very expensive as
ure are as follow: it is iterative in nature. As a solution I precom-

4
pute a distance matrix K. The entry Mi, j is the • Plan points are chosen to maximize the in-
geodesic distance between the points i and j on formation gain.
the mesh. Using this strategy the computation
time for Pdist is greatly reduced. Computation • Entropy value is used to measure the in-
of Pdist can be formation content.

Pdescr ∼ exp(−0.5σd−2 | Dx − Dy |)(1)


7 Results:

The entropy based algorithm has shown


promising results. Planned sampling strategy
Pdist (x, y| x1 , y1 , ..., xk , yk )∼ adds to the efficiency in solving the problem of
∏ki=1 exp(−0.5σ −2 g (dMs (x, xi ) − shape matching. The maximum accuracy that
2
dMt (y, yi )) )(2) I reached on this bunny data-set is 42 percent.
This number can be increased in subsequent ex-
periments by sampling more number of points
Psamp (x, y | x1 , y1 , ..., xk , yk ) = on both the models. Bunny model is a tough
Pdesc (x, y).Pdist (x, y | x1 , y1 , ..., xk , yk )(3) one for this experiment. The descriptors used
for this algorithm are very same in highly pla-
nar regions. This has also caused a difficulty
Random Matching algorithm is as follows: in matching. If point chosen at random lies on
a flat region then its descriptor is similar to a
• For first point: Randomly choose x, then
large number of other points. Hence If the ini-
choose y with the maximum Pdescr .
tial guesses are good then there is more chance
• Given we already have k correspondences of better final result.
(x1, y1),.., (xk, yk), (xi, yi) MSMT, the
likelihood for one more correspondence
(x, y) MSMT being correct is given by
a Gaussian error model Pdist .

• Pdist is the likelihood that for a new ran-


dom pair [x, y] the intrinsic distances dis-
tances for all the previous matches is con-
served.

• For rest of the points: Randomly choose


x, then choose y with the maximum
Psamp Sampling is continued until no more
matching pairs are found that do not ex-
ceed the maximum error threshold or until
all points are matched.
Result from the paper: First column shows
Planned Matching Algorithm: resampled points using poisson disk sampling.
Other images represent the point correspon-
• Uses Plan points for Random Matching al- dences found using the landmark sampling
gorithm. technique

5
References
[1] A NGUELOV, D., S RINIVASAN , P., PANG , H.-C.,
KOLLER , D., T HRUN , S., AND DAVIS , J. The cor-
related correspondence algorithm for unsupervised
registration of nonrigid surfaces. In Advances in
Neural Information Processing Systems. MIT Press,
2005, pp. 441–448.
[2] B ESL , P. J., AND M C K AY, N. D. A method for
registration of 3-d shapes. IEEE Trans. Pattern Anal.
Mach. Intell. 14, 2 (1992), 239–256.
[3] C HEN , Y., AND M EDIONI , G. Object modelling by
registration of multiple range images. Image Vision
Comput. 10, 3 (1991), 145–1556.
[4] G ELFAND , N., M ITRA , N. J., G UIBAS , L. J., AND
P OTTMANN , H. Robust global registration. In Pro-
ceedings of the Third Eurographics Symposium on
Geometry Processing (2005), Eurographics Associ-
ation.
[5] H UANG , Q.-X., A DAMS , B., W ICKE , M., AND
G UIBAS , L. J. Non-rigid registration under isomet-
ric deformations. In Proceedings of the Symposium
on Geometry Processing (2008), Eurographics As-
sociation, pp. 1449–1457.
[6] J.S.B. M ITCHELL , D. M. M., AND C OMPUT.,
C. H. P. S. J. Multiple source/target exact geodesic
(shortest path) algorithm for triangular mesh (trian-
gulated 2d surface in 3d).
[7] M ITRA , N. J., G ELFAND , N., P OTTMANN , H.,
AND G UIBAS , L. Registration of point cloud
data from a geometric optimization perspective.
In Proceedings of the 2004 Eurographics/ACM
SIGGRAPH Symposium on Geometry Processing
(2004), ACM, pp. 22–31.
[8] RUGGERI , M. R., AND S AUPE , D. Isometry-
invariant matching of point set surfaces. In Proceed-
ings of the 1st Eurographics Conference on 3D Ob-
ject Retrieval (2008), pp. 17–24.
[9] T EVS , A., B ERNER , A., WAND , M., I HRKE , I.,
AND S EIDEL , H.-P. Intrinsic shape matching by
planned landmark sampling. Eurographics (2011).

You might also like