Registration and Integration of Textured 3-D Data

Sing Bing Kang

Cambridge Research Laboratory
Digital Equipment Corporation
sbk @ crl.dec.com

Andrew Edie Johnson

The Robotics Institute
Carnegie Mellon University
aej @ri.cmu.edu

Abstract
In general, multiple views are required to create a
complete 3-0 model of an object or of a multi-roomed
indoor scene. In this work, we address the problem of
merging multiple textured 3-0 data sets, each of which
corresponds to a different view of a scene or object. There
are two steps to the merging process: registration and
integration. To registel; or align, data sets we use a modiJied
version of the Iterative Closest Point algorithm; our
version, which we call color ICE considers not only 3-0
information, but color as well. We show that the use of color
decreases registration error by an order of magnitude. Once
the 3 - 0 data sets have been registered, we integrate them to
produce a seamless, composite 3-0 textured model. Our
approach to integration uses a 3 - 0 occupancy grid to
represent likelihood of spatial occupancy through voting. In
addition to occupancy information, we store surjiuce normal
in each voxel of the occupancy grid. Suqace normal is used
to robustly extract a surj-iace f m m the occupancy grid; on
that surjiace we blend textures from multiple views. .

1. Introduction
There is an increasing interest in modeling scenes for
virtual reality applications, either in the areas of business
(real estate, architecture, information-dispensing kiosks,)
education (electronic museums and multimedia books), or
entertainment (interactive 3-D games, movies). The option
of creating virtual environments by capturing real scenes
through video cameras is receiving particular attention,
given the labor-intensive and thus expensive nature of
creating models by hand using a 3-D geometric modeler.
The problem with creating models of a large and complex
scene is that a single view cannot completely convey the
shape of the scene--thus merging of multiple views acquired
at different locations is usually necessary. I n general,
merging of multiple views is a two step process: first, the
views are registered, then they are integrated into a seamless
3-D model.

Real scenes are described by the shapes of objects in the

scene as well as by the appearance (color, texture) of these
objects. Therefore, for computer models to be realistic, they
must convey both shape and appearance. To meet this end,
we have developed a volumetric multi-view merging
algorithm that integrates the shape and appearance of
complex scenes.
There has been much research in the area of model
creation through multiple view merging. Shum et. al. [15],
for example, recover the merged model through
simultaneous determination of planar surface parameters
and pose of constituent range data sets. They assume,
however, that the surfaces of objects can be represented
using planar patches. There have been many volumetric
approaches to 3-D data integration [2][5][6][
171, but none
of them were designed to merge texture as well as shape.
Furthermore, these algorithms assume that the views have
already been aligned. Our approach builds on these
approaches to solve the problem of merging shape and
texture for generation of more descriptive scene models; our
algorithm also includes a method for accurately registering
textured 3-D data sets.

2. Recovery of 3-D scene data

In our work, we use 3-D data recovered from
omnidirectional multibaseline stereo, i.e., using multiple
panoramic images [8]. Each panoramic image spans a 360"
horizontal field of view. The primary advantage of this
method is that, at any given camera center location, the
scene can be recovered at a very wide horizontal field of
Omnidirectional
Multibaseline
Stereo

Textured
3-D Data

Figure 1. Shape recovery using omnidirectional

multibaseline stereo.

rhis work was done at the Cambridge Research Laboratory and wa5
supported by Digital Equipment Corporation.

0-8186-7943-3/97 $10.00 0 1997 IEEE

Recovered
Points

234

view. This is done without resorting to any intermediate 3D merging.

The omnidirectional multibaseline stereo approach to
recover 3-D data and, subsequently, the scene model is
summarized in Figure 1. We provide only a brief outline of
the approach here; full details can be found in [8]. The
approach is straightforward: at each camera location in the
scene, sequences of images are captured while the camera
is rotated about the vertical axis passing through the camera
optical center. Each set of images is then composited to
produce panoramas at each camera location. The stereo
algorithm is then used to extract 3-D data of the scene. A
surface represented as a triangular surface mesh is
constructed from the 3-D data. Finally, a textured 3-D data
set results when the surface mesh is rendered with the
texture provided by the reference 2-D input image.
Given multiple textured data sets recovered using
omnidirectional multibaseline stereo, the first step in the 3D merging process is data set registration.

the texture projected from all of the data sets must be

accurately aligned on the final consensus surface. In other
words, for correct alignment of texture, registration error on
the order of a few image pixels projected into the scene is
required. For example, a 2000 pixel wide panorama
becomes misregisteredby one pixel if the estimated rotation
is incorrect by 0.18 degrees. Inaccuracies in scene shape
introduced by
the
shape recovery
algorithm
(omnidirectional stereo [SI) are too large to obtain the
accuracy in registration needed to blend texture using a
traditional ICP algorithm. However, by including color in
the closest point computation of the ICP algorithm, the
necessary registration accuracy can be obtained. This is
accomplished by modifying the distance metric used in the
closest point computation between points p1 =
(XI 1tx12,x13>cl19c12,c13) and p2 = (x2 I ,x223x23,c21~ ~ 2 2 , ~ 2 3 )
to include both shape and color information (xs and cs
respectively), i.e.,

3. Registration

d6(pl,p2) =

The technique that we use to register all the 3-D data sets
is essentially a modification of the Iterative Closest Point
algorithm (ICP) [l]. In addition to using k-d trees for
efficient closest point computations and a dynamic distance
threshold [18], our algorithm uses shape and color
information to improve the registration beyond that
obtained with an ICP algorithm that uses just shape
information. We call this variant the Color ICP technique;
its concept is shown in Figure 2.
During integration of textured 3-D data, shape as well as
texture are combined to form what is called the final
consensus surface model. Our approach to texture
integration is to project the texture from all of the registered
data sets onto the final consensus surface where the
overlapping textures are blended.
For our algorithm to be able to blend texture correctly,
Traditional ICP
Before
Registration

After
Registration

closest
points

tt

incorrect texture
alignment

r2(c,2

- c22)

2
+rdC13

-9 3 )

(l)

where r= (Tl,r2,r3) are scale factors that weigh the

importance of color against the importance of shape. In
Color ICP, a six dimensional k-D tree is used to speed up
closest point computations.
A measure of registration error is the distance between a
point after registration and the true location of the point. A
histogram of this distance, for all of the points in a synthetic
textured 3-D data set (where the correct registration is
known) is shown for the traditional ICP and Color ICP
algorithm in Figure 3. It illustrates the order of magnitude
improvement in registration using color information in
addition to 3-D shape, making integration of texture
feasible.

Color ICP

points

( x ~ ~ -2 +
x r~1 (~c ,) I - ~ 2 , +)

600 I

v\v

Histogram of Registration Error

+d

--tI)-cf-

I3

8
correct texture
alignment

Figure 2. Demonstration of the use of color in

registration. In traditional ICP closest points
depend only on shape, so incorrect texture
registration can occur. Since closest points
depend on color and shape in Color ICP, it will
align texture correctly.

Log10 Distance of Point from Correct Position

Figure 3. Histogram of registration errors for the

traditional ICP and Color ICP algorithms. The
histogram clearly shows that the Color ICP
algorithm is an order of magnitude improvement
over the traditional ICP algorithm.

235

geometry is shown in Figure 4.

Matthies and Shafer [12] showed that the variances of
the sensor crror model should vary depending on the
position of the sensed point. To reduce the amount of
calculation per point, we have assumed that the variances of
the sensor error model are fixed for all points. However, the
variances of the model can be automatically set by
analyzing local changes in distance from the sensor which
characterize the noise in the recovery of the 3-D points.
Consider a pointp from surface mesh M that has NM points
and sensor origin S . Call the local surface mesh
neighborhood o f p (points connected top by the mesh), L,
with Np points. The RMS spread in distance d, for M is
calculated as follows:

4. Integration
After registration, the data sets are integrated to produce
a seamless 3-D textured model. The integration step
involves the use of occupancy grids based on sensor
modeling and robust ridge detection to recover composite 3D surfaces. Occupancy grids [3][ 1 13 represent the
likelihood of 3-D spatial occupancy through voting. This
representation is very attractive for integrating 3-D textured
data sets because it is incremental, simple, allows for freeform objects, and flexible in allowing the incorporation of
data from different sensors and their models.

4.1 Sensor model

Before accumulating surface evidence in an occupancy
grid, a sensor model must be determined. Our sensor model
G has two components: the sensor error model and the point
spread model. The sensor error model GE is an
approximation of a true stereo error model and is
represented as an ellipsoidal gaussian distribution centered
at the measured 3-D point whose axis is oriented along the
line of sight [12]. The point spread model Gs is used to
promote the generation of surfaces from discrete data. It is
represented as a cylindrical gaussian, centered at the sensed
point, whose axis is aligned with the local surface normal.
A linear combination is used to combine the sensor error
and point spread models into one sensor model G.

d,
measures the average local change in distance along
the viewing direction which is a good measure of sensor
error assuming that neighborhoods are locally planar, with
normals roughly oriented along the viewing direction. The
variances in the sensor error model are set automatically
based on estimated error as oa = d,,,y and op = 2d,,,7.
Stereo returns discrete point measurements on the
surface of objects. By spreading the contribution of a point
along the tangent plane of the point, a continuous surface
can be generated. To meet this end, a point spread model is
added to the sensor model. The point spread model has the
form of a cylindrically symmetric gaussian with its axis
aligned with the local surface normal

By adjusting the parameter h on the interval [0,1], the

relative importance of the sensor error and point spread
models can be set. Convolution of the point spread model
with the sensor error model is a more rigorous way of
combining the models, but computationally we found it
infeasible because both models change dynamically with
the point being processed.
Analytically, the sensor error model G, has the form of
a cylindrically symmetric gaussian with its axis aligned
with the local viewing direction

Y = */r.x-x.fi
6 = x.2
x = Q-P
where g is the distance of the query point x from the unit

where a is the distance of the query point x from the unit

viewing vector P and p is the distance of the query point x
along the unit viewing vector. S , P and Q are the world
coordinates of the sensor, the data point, and an arbitrary
point, respectively. The spread of the gaussian can be
characterized by two parameters, CY,* the variance
perpendicular to the viewing direction and G$ the variance
along the viewing direction. A 2-D slice of the sensor error

Sensor Error Model GE

Point Spread Model GS

Figure 4. Geometry for the two components of the

sensor model: the Sensor Error Model, a
cylindrical gaussian oriented along the sensor
viewing direction and the Point Spread Model, a
cylindrical gaussian oriented along the surface
normal.

236

surface normal n and d is the distance of the query point x

along the unit surface normal. The spread of the gaussian
can be characterized by two parameters, 0: the variance
2
along the tangent plane and od the variance along the
surface normal. A 2-D slice of the surface spreading
geometry is given in Figure 4.
The variances of the point spread function can be
calculated automatically for each surface mesh by
estimating the local resolution at each point. Ideally the
variances of the spread function would be different for each
point in the surface mesh, since the local resolution changes
for each point. However, to reduce the computation for each
point, the variances are fixed for each surface mesh and are
based on the average mesh resolution for all of the points in
the mesh. The average mesh resolution r,, for a surface
mesh M with N E edges is

More specifically, the gradient of surface likelihood is

computed using finite differences at each voxel of the
occupancy grid. Next, the dot product of the surface
likelihood gradient (g) and the consensus surface normal
(n)are computed at each voxel. As shown in Figure 5, this
dot product defines an implicit surface function ( I ) which
will be positive on one side of the ridge and negative on the
other side. This ridge detection method is more robust than
traditional ridge operators because the direction along
which to calculate the local maxima for ridge detection is
already given by the consensus surface normal. Usually
ridge detection requires computation of second order
surface derivatives to determine this direction--a
computation without a robust solution.
The implicit surface function is then polygonized using
the standard Marching Cubes algorithm [IO] with a
modified lookup table of 22 cases to prevent the creation of
holes in the surface [13]. The consensus surface mesh
generated from the 6 synthetic data sets is shown in Figure
7. Once the consensus surface has been recovered, the next
step is to generate its texture.

Based on the average mesh resolution, the variances of the

point spread function can be set as oY = 2ruv and

c8 = r a y .

4.3 Blending texture

An example of the recovered surface probability

distribution is shown in Figure 7(b) for six registered data
sets. Brighter values correspond to higher probability.

Texture blending is done by weighted averaging of

overlapping textures from the original contributing data
sets. The texture weight is a function of the angle between
the surface normal and the viewing direction; textures that
subtend more than 90" with the consensus surface normal
are discarded.
Suppose there are N textured 3-D data sets to be
integrated. For texture blending, we store an additional

4.2 Recovering consensus surface

While the traditional occupancy grid stores only the
likelihood of occupancy, we encode the consensus surface
normal and surface likelihood in a 3-vector in each voxel.
The magnitude of the vector corresponds to surface
likelihood and the direction corresponds to the consensus
surface normal (likely local surface normal). As shown in
Figure 5, vector addition instead of scalar addition is used
to update the occupancy grid. When a new sensor
measurement is inserted into a voxel, a vector oriented in
the direction of the measurement surface normal with
magnitude equal to the sensor model at that voxel is added
to the vector already stored in the voxel. Therefore, each
voxel contains a weighted sum of the surface normals of
nearby data points; this gives an estimate of the most likely
surface normal at that point in space given the sensed data.
Using a vector representation improves the surface
generation because it provides the most likely local surface
normal, enforces shape coherence and prevents mixing of
spatially close, but opposing surfaces which is particularly
important when integrating texture. The benefits of the
vector representation are demonstrated in Figure 6.
The surface of the merged data set is recovered by
detecting ridges in surface likelihood. In our case, the ridge
can be computed as the local maxima of the surface
likelihood in the direction of the consensus surface normal.

consensus
surface
normal

$. $.

"

likelihood
gradient

Figure 5. The occupancy grid stores consensus

surface normal and is updated by each data point
using vector addition (top). The consensus
surface normal defines the direction along which
to search for local maxima in surface likelihood.
(bottom).

237

Coherence

Opposing Surfaces

Consensus Surface Normal

surface

3
surface

high probability
on surface
low probability
betweensurfaces
high probability
on surface

+ + +
=
coherent = high probability

partially coherent = average probability

not coherent = zero probability

Figure 6. Consensus surface normal definitions. The consensus surface normal is the weighted sum of
the normals of surrounding points (left). Adding probabilities as vectors prevents opposing surfaces from
mixing (middle). Coherence of normals determines magnitude of consensus surface normal (right).

(d)

(e)

Figure 7. Registered points sets (a) with sensor origins shown as shaded spheres and the middle
horizontal slice of surface probability through the voxel space for those points (b). Notice that only
allocated voxels are shown. Three views of the consensus surface mesh generated for six registered data
sets are also shown (c)-(e).

Figure 8. Two representative panoramas of the vision lab.

238

vector in each voxel for each data set being integrated; this
vector encodes the vector contribution of the data set to the
consensus surface normal of that voxel. Suppose there are N
vectors ni in each voxel, and those vectors measure the
contribution of each data set i to the consensus surface
normal n, of the voxel. The texture weight w iof each data
set is then the dot product of the consensus surface normal
with the contribution of that data set to the voxel
w, = man(0, n i . n,) .
(7)

To apply texture to the consensus surface mesh, a small,

square texture map, called a texture cell, is made for each
face. The color of the pixels of the texture cell are then
determined by finding the 3-D position of the pixels on the
plane of the face followed by trilinear texture blending at
the point. Alternatively, the texture at a pixel in a cell can be
set to the texture projected from the data set with maximum
texture weight. The geometry of texture blending is
illustrated in Figure 9. Figure 10 shows the result of texture
blending six textured 3-D data sets.
We have also worked with real scenes; results of one
experiment with a real office scene are shown in Figure 1I.
Figure 12 and Figure 13 show the results for merging
two other real data sets, this time from two views of our
vision lab. These two data sets mostly intersect, except that
the first data set includes the small back room in the lab
while the other data set does not. The reference panoramic
images corresponding to the two data sets are shown in
Figure 8.
A difficulty with these data sets stems from the fact that
the door to the back room (with the stack of VCRs) is
relatively narrow, causing the algorithm that creates the 3-D
mesh to connect across the doorway for the second data set,
as shown at the bottom left of Figure 12. As a result, a
preprocessing step that culls points that violate visibility of
other data points is performed; the results of this step is
shown in Figure 12. The results of merging the two data sets
are shown in Figure 13. The discontinuity in the resulting
combined texture bears testimony to the fact that the
recovered shapes at the two different sites are not exact.

If wi is negative then ni is pointing away from the

consensus surface normal. This means that the sensor origin
of data set i is on the opposite side of the consensus surface,
so data set i should not be contributing texture to the
surface. Therefore, if w iis negative, it is set to zero. Using
the dot product to create texture weights is the same as
setting the texture weight equal to the ratio of area visible to
the sensor to actual surface area. This is a reasonable
heuristic for vision sensors because as the ratio of visible to
actual surface decreases, the reliability of the appearance
measured decreases. Furthermore, we are eliminating the
need for ray-tracing by storing the relative contribution of
each data set in each voxel.
Because the consensus surface mesh was created using
Marching Cubes, each face in the surface mesh lies in a
cube formed by eight voxels. A simple method for
determining the texture weights at a point P on a face in the
cube is to trilinearly interpolate the texture weights from the
eight voxels based on the 3-D location of P in the cube.
Then the color at P is the texture weighted average of the
color projected onto P from each data sets.Since trilinear
interpolation of image weights is used, the texture will vary
continuously over the faces.

Textured Surface

Texture Close-up

= w c +wc + w c
1 1
2 2
3 3

Figure 9. The geometry for creating a texture map

cell for a face. The color of each pixel of the cell is
the weighted average of the colors projected onto
it by data sets that view the pixel. Each face in the
consensus surface mesh has an associated
texture cell.

Figure 10. The result of integrating six textured 3D data sets created directly from a synthetic room
model. The complete room model with texture
blended on the surfaces of the room is shown as
well as a close up of the texture blending.

239

break out of local minima.

One of the problems associated with the integration step
is the sensitivity of the results of texture blending to the
accuracy of the recovered shape. There is very little
recourse to bad input data, although a more sophisticated
structure from motion algorithm may be bootstrapped to the
registration step to improve both relative camera pose and
3-D data.
The work described here is used to recover 3-D models
of indoor scenes for the on-going Smart Kiosk project at
Cambridge Research Lab., Digital Equipment Corp. [ 161.
The Smart Kiosk can be described as an enhanced version
of the Automatic Teller Machine, with the added capability
of being able to interact with the user through body
tracking, and gesture and speech recognition. The recovered
model of the environment would allow the kiosk to situate
the user relative to the environment. As a result, it would
enable a more engaging level of user-kiosk interaction,
specifically being able to provide relative directions as well
as give a virtual tour of the environment. The incorporation
of this enhanced feature (using the recovered model of the
environment) to the Smart Kiosk is currently underway.
It is perhaps worthwhile to investigate an alternative,
view interpolation-based means of generating synthetic
views of the model. However, these methods are not
appropriate whenever 3-D structural information of the
scene is desired or when certain kinds of views (such as flythroughs involving camera positions very different than
those of the known sampled views) are desired.

5. Implementation
The code for our model merging work is written in C++
and uses the Library of Efficient Data types and Algorithms
(LEDA) [14]. LEDA is a library of data types and
algorithms that includes, among others, graph data
structures and algorithms to manipulate them. Each vertex,
edge, and face of a 3-D scene model has its own data
structure, while the connectivity information between the
vertices is encoded in a graph. This graph represents the
geometrical surface mesh of the 3-D model. Meanwhile, the
occupancy grid is represented as a dynamically allocated
list of voxel structures; each voxel structure contains the
surface normal and probability information. By
implementing the allocated voxels as a dictionary, the
access to the voxel structure is made efficient. The 3-D data
merging and modeling program is compiled and run on a
DEC UNIX Alpha workstation.
While we have written our own version of a 3-D model
viewer, we also provide a facility to output our 3-D models
in VRML. Some of the results from this paper can be
viewed on web at http://www. ius.cs.cmu.edu/usr/users/aej/
www/research/vrml-modeling.
html using a VRML browser.

6. Discussion and future work

It is not surprising that adding color information to the
registration step improves performance. There is a danger,
on the other hand, of adding many more local minima with
color. This is clearly a function of both the shape and the
texture distribution. Repetitive shape and texture would
have an adverse influence. A solution to this may be to add
a simulated annealing-like characteristic to the algorithm to
Points (top
. . view)

7. Summary
We have described our approach to merging multiple

Surface (oblique
view)
.
.

Original Set #1

Processed Set #1

Original Set #2

Processed Set #2

Textured Surface (oblique view)

,
Figure 11. The result of integrating two textured 3D data sets created with omnidirectional stereo of
an office. The registered points, wire frame
surface, texture mapped surface and close-up of
the texture mapping are shown. Note: the two
small spheres in the top view of the 3-D points
represent the different camera center locations.

Figure 12. The result of culling of two data sets of

the vision lab based on visibility. Left: original
surface meshes, right: culled surface meshes.

240

textured 3-0data sets. In our work, the 3-D data sets are
recovered using omnidirectional multibaseline stereo,
which involves multiple panoramic images of the scene.
Data merging is a two-step process, namely registration
and integration. I n registering multiple data sets using a
variant of the ICP algorithm called the Color ZCP, w e
consider not only 3-D point location, but also color
information. The color information has been shown to
improve the registration significantly, especially if there is
ambiguity in using only 3-D information.
Once the multiple data sets have been registered, we then
extract the complete model. The construction of the merged
model is based o n voting through occupancy as well as
consistency in surface normal direction. T h e surface of the
merged model is recovered by detecting ridges in the
occupancy grid, and subsequently polygonized using the
standard Marching Cubes algorithm. The texture o n the
complete model is determined through trilinear
interpolation of the overlapping textures corresponding to
the original data sets.
A more detailed description of our work on merging
different textured 3-D data sets can be found in [7].
Points (top view)

References
[l] P. Besl and N. McKay. A method of registration of 3-D
shapes. IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 12, no. 2, pp. 239-256, February 1992.
[2] B. Curless and M. Levoy. A volumetric method for building
complex models from range images. Computer Graphics
(SIGGRAPH 96).August 1996.
[3] A. Elfies. Sonar-based real world mapping and navigation.
IEEE Joul: Robotics and Automation, vol. RA-3, no. 3, pp,
249-265, 1987.
[4] K. Higuchi, M. Hebert and K. Ikeuchi. Building 3-D models
from unregistered range images. Technical Report CMU-CS93-214, Carnegie Mellon University,November 1993.
[5] A Hilton, A. Stoddart, J. Illingworth and T Windeatt.
Reliable surface reconstruction from multiple range images.
Fourth European Con$ on Computer Vision (ECCV 96),pp.
14-18, April 1996.
[6] H. Hoppe, T. DeRose, T. DuChamp, J. McDonald and W.
Stuetzle. Surface reconstruction from unorganized points.
Computer Graphics (SIGGRAPH 92),pp. 71-78. July 1992.
[7] A. Johnson and S. Kang. Registration and integration of
textured 3-D data. Digital Equipment Corp. Cambridge
Research Lab. Tech. Report CRL 96/4, October 1996.
[8] S. Kang, and R. Szeliski. 3-D scene data recovery using
omnidirectional multibaseline stereo. Digital Equipment
Corporation Cambridge Research Lab. Tech. Report CRL
95/6, October 1995.
[9] C. Kolb. Rayshade User Guide and Reference Manual.
August 1994.
[lo] W. Lorensen and H. Cline. Marching Cubes: a high
resolution 3D surface construction algorithm. Computer
Graphics (SIGGRAPH 87),pp. 163-169, 1987.
[l 11 M. Martin and H. Moravec. Robot Evidence Grids. Carnegie
Mellon University Robotics Institute Tech. Report CMU-RITR-96-06,March 1996.
121 L. Matthies and S. Shafer. Error modeling in stereo
navigation. IEEE Jour Robotics and Automation, vol. RA-3,
no. 3, pp. 239-248, June 1987.
131 C. Montani, R. Scateni and R. Scopigno. A modified look-up
table for implicit disambiguation of Marching Cubes. Visual
Computer, vol. 10, pp. 353-355, 1994.
141 S. Naher and C. Uhrig. The LEDA User Manual. May 1996.
151 H.-Y. Shum, K. Ikeuchi and R. Reddy. Principal component
analysis with missing data and ils application to object
modeling. Proc. IEEE Computer Vision and Pattern
Recognition (CVPR-941, pp. 560-565, June 1994.
[16] K. Waters. J. Rehg, M. Loughlin, S. Kang and D.
Terzopoulos. Visual sensing for humans for active public
interfaces. Cambridge, UK, 1996
[ 171 M. Wheeler. Automatic modeling and localization for object
recognition. Carnegie Mellon Univ., School of Computer
Science Tech. Report CMU-CS-96-118,October 1996.
[18] Z. Zhang. Iterative point matching for registration of freeform curves and surfaces. Intl Jour: Computer Vision, vol.
13 no. 2, pp. 119-152, 1994.

Surface (oblique view)

.<

Surface (top view)

Textured Surface (oblique view]

Close-uD #1 Imax-blend)

Figure 13. the result of merging two textured 3-D

data sets created with omnidirectional multibaseline stereo of a lab. The texture of the merged
model is created using the max texture blending
scheme.

241