Tomas Pajdla
Computer Vision Laboratory
Czech Technical University
Karlovo nam. 13
CZ 121 35 Praha 2
pajdla@vision.felk.cvut.cz
Abstract
I will present a technique for camera calibration and Euclidean reconstruction from
multiple images of the same scene. Unlike standard Tsai's camera calibration from
known scene we exploit controlled known motions of the camera under scrutiny to
arrive at the calibration and Euclidean reconstruction without any knowledge about
the scene.
The use of special motions, in this case pure translations, allows us to obtain a
linear solution which is, for linear camera, more stable than Tsai's "bundle adjust-
ment". In our case only 8 (5 internal and 3 external) parameters are to be estimated
comparing to more than 11 parameters which are to be found in the standard case.
We consider three translations of an uncalibrated but same camera providing
us with four views of the scene. We also assume to measure the translation vectors
in some Euclidean coordinate system rigidly attached to the camera, yet in an
unknown relation to the camera affine coordinates. This, for instance, can easily
be accomplished by mounting the camera on a robot's arm. Moreover, we assume
to track at least four points in all four views. Such special, but still realistic,
arrangement bring us a linear relation C x = 0, where the matrix C is a function
of known translations as well as image coordinates of tracked points. The vector
x contains all intrinsic camera calibration parameters, rotation parameters of the
camera with respect to the robot's coordinate system (this solves part of the eye-
hand orientation problem), and proper scaling factors for all points allowing their
Euclidean reconstruction.
The experiments show that by exploiting Total Least Squares in combination
with careful normalization of image coordinates, an efficient algorithm outperform-
ing lengthy and vulnerable nonlinear optimization can be obtained.
Jan Flusser and Tomas Suk
Institute of Information Theory and Automation
Academy of Sciences of the Czech Republic
Pod vodarenskou vezi 4
182 08 Prague 8, Czech Republic
e-mail: flusser@utia.cas.cz
Abstract
The paper will be devoted to the feature-based recognition of blurred images ac-
quired by linear shift-invariant imaging system against an image database.
In such system, the imaging process is described by a convolution
g(x; y) = (f * h)(x; y) (1)
where f (x; y) and g(x; y) represent the original and observed images, respectively
and h(x; y) is a system point spread function.
The proposed approach consists of describing images by features which are in-
variant with respect to blur (that means with respect to the PSF) and recognizing
images in the feature space. In comparison with complicated and time-consuming
"blind-restoration" approach, we do not need the PSF identification and image
restoration. Thanks to this, our approach is much more effective.
Index Terms: image recognition, linear imaging system, symmetric blur, motion
blur, image moments, blur invariants
Margaret M. Fleck
Department of Computer Science
University of Iowa
Iowa City, IA 52242 USA
mfleck@cs.uiowa.edu
Abstract
A human observer sees about 180 degrees horizontally and 120 degrees vertically.
A single glance supplies this field of view at low resolution; fast eye and head
movements can bring any part of it into high resolution. We use a wide field of
view to see the overall shape of large objects and objects in confined spaces, to
robustly estimate egomotion, to locate objects and detect hazards in our immediate
environment, and to estimate global properties of our location (e.g. time of day,
geographic location). For related reasons, wide-angle lenses are frequently used in
surveillance, nature, and news photography.
Computer vision research has concentrated on images subtending no more than
60 degrees. The perspective imaging model assumed by most algorithms breaks
down for wide-angle images. Past about 80 degrees, the shape of peripheral objects
is significantly distorted and C-mount lenses display significant barrel distortion.
For fields of view more than 120-140 degrees, size distortion becomes so extreme
that perspective lenses can no longer be designed. At 180 degrees, the model fails
completely. One available (though expensive) 35mm lens has a 220 degree field of
view: obviously it is not designed to a perspective model.
Avoiding these problems with perspective requires two changes in method. First,
the ideal imaging model should be one of the fisheye projections used in wide-angle
lens design. Three of these projection models_equidistant, stereographic, and equi-
solid angle_can handle very wide fields of view. Stereographic projection is the
most convenient mathematically: circularity, intersection angles, and local shape
are preserved by the mapping from a spherical to a stereographic image. As a
consequence, local symmetries of 3D objects project onto local symmetries of the
2D outline whenever simple visibility conditions are satisfied.
Second, camera calibration or self-calibration is a necessity for most wide-angle
algorithms. Wide-angle lenses diverge not only from perspective projection but from
one another. Standard calibration algorithms (e.g. Tsai's algorithm) are complex
and their models of radial distortion are inadequate for wide-angle lenses. We have
developed a simple technique for calibrating wide-angle lenses, using controlled
camera motion.
Using a camera calibration, images can be transformed into ideal stereographic
or spherical projection. In spherical projection, the apparent shape of a 3D object
no longer depends on its location in the field of view, but only on the tilt of the
object relative to the observer. A similar approximate result holds for stereographic
images, as long as the object subtends a small visual angle. This simplifies object
recognition.
Fridrich Sloboda and Bedrich Zatko
Institute of Control Theory and Robotics
Slovak Academy of Sciences
Dubravska 9, 842 37 Bratislava
Slovak Republic
e-mail: utrrzatk@savba.sk
Abstract
Piecewise linear approximation of planar Jordan curves and arcs by gridding tech-
niques is described. The approximation is based on the basic notions of intrinsic
geometry of metric spaces: on the notion of the shortest path in a polygonally
bounded compact set and on the notion of geodesic diameter of a polygon. Conver-
gence of piecewise linear approximations to rectifiable planar Jordan curves and arcs
is investigated. Comparison of piecewise linear approximations on orthogonal and
triangular grids is performed. Algorithms for the shortest path problem solution
in a polygonally bounded set and for geodesic diameter calculation are described.
Applications are highlighted.
J. Matas, R. Marik, and J. Kittler
Vision, Speach, and Signal Processing Group
University of Surrey, Guilford
United Kingdom
matas@ee.surrey.ac.uk
Abstract
A new representation for objects with multiple colours - the colour adjacency graph
(CAG) - is proposed. Each node of the CAG represents a single chromatic com-
ponent of the image defined as a set of pixels forming a unimodal cluster in the
chromatic scattergram. Edges encode information about adjacency of colour com-
ponents and their reflectance ratio.
The CAG is related to both the histogram and region adjacency graph repre-
sentations. It is shown to be preserving and combining the best features of these
two approaches while avoiding their drawbacks. The proposed approach is tested
on a range of difficult object recognition and localisation problems involving com-
plex imagery of non rigid 3D objects under varied viewing conditions with excellent
results.
Radima Sara
Computer Vision Laboratory
Faculty of Electrical Engineering
Czech Technical University
sara@vision.felk.cvut.cz
Abstract
We study formation of intensity image of an opaque surface and we consider the
question of how do the structure and the properties of the 3-D curvilinear surface
project on the structure and the properties of the perceived image. The talk will
be focused on this topic restricted to static monocular intensity images.
The occluding contour, the self-shadow contour, the image singularities, the
zero-crossings of image brightness iso-levels curvature, the zero-crossings of the
image gradient curves curvature, and the shading at a general image point will be
examined. It will be shown that
o there is a class of surfaces whose shading does not contain sufficient informa-
tion for recovering any of the surface and the light parameters,
o there are structures in image whose certain properties are invariant to the
reflectance function,
o there are curves in image along which the surface normal and the surface qual-
ity (the sign of Gaussian curvature) can uniquely be determined independent
on illuminant position, the surface texture, and the reflectance function,
o there are easily detectable isolated points in image whose pre-images are lo-
cated at surface parabolic lines (i.e., the lines of vanishing Gaussian curva-
ture),
o there are isolated points in image where the tilt component of the light direc-
tion vector can uniquely be determined independent on the surface texture
and the reflectance function,
o there are curves in image in the vicinity of which the surface interpretation
(convex/concave/elliptic/hyperbolic) can only change (assuming Lambertian
surface without texture),
o there are isolated points in image where the square of the shape operator ma-
trix can be determined if the reflectance function is Lambertian, the imaging
geometry is known, and the surface is not textured; this means that up to
four local surface interpretations are possible there,
o at a general image point, the local surface interpretation (the normal vector
and the shape operator) is ambiguous; up to eight interpretations are possible
even if the surface is not textured, and the reflectance function and the imaging
geometry are known.
References
[1]R. Sara. What can and what cannot be learned about 3-D shape from a single
intensity image? In Proc. Vision Milestones 1995, to be published in Springer
Verlag.
[2]R. Sara. Isophotes_the key to tractable local shading analysis. In Proc.
CAIP '95, 1995. To be published.
[3]R. Sara. Local Shading Analysis via Isophotes Properties. PhD thesis, Jo-
hannes Kepler University Linz, Dept. of System Sciences, March 1994.
Mourad Zerroug and Gerard Medioni
Institute for Robotics and Intelligent Systems
University of Southern California
Los Angeles, CA 90089-0273
medioni@tiresias.ucs.edu
Abstract
Recognizing 3-D objects in an image is a complex problem along multiple dimen-
sions. One has to extract relevant "descriptions" from the image, "match" them
with one or more of the objects from a previously constructed database, "learn" new
objects for inclusion in the database. While these tasks are necessary, no consensus
has emerged regarding the choice and level of features (2- D or 3-D), the type of
indexing used, the matching strategy, and the order in which these tasks should be
performed.
Most existing recognition systems use low-level representations such as points,
lines and collections of these. Since these descriptions only carry minimal seman-
tic contents, the indexing scheme cannot be very efficient, and the computational
burden falls on the localization phase which uses geometric constraints. This can
only succeed when the database consists of exact models. Furthermore, learning
can only occur with user supervision. Instead, we argue that the only way to solve
all the previous tasks is to generate and use high-level representations in terms of
volumetric parts, their hierarchy and their arrangement. We further propose that
part-based representations based on subsets of Generalized Cylinders constitute
suitable means for shape abstraction of large classes of complex 3-D objects, and
that these descriptions can indeed be generated from 2-D images under imperfect
and realistic viewing conditions.
We discuss here the challenging sub-tasks that need to be accomplished in order
to build systems capable of describing complex objects from monocular intensity
images, and the elements of a recognition engine to manipulate and reason about
the resulting hierarchical descriptions. We first discuss the role of the representa-
tion scheme itself in providing generic constraints for the extraction of part-based
descriptions. We argue that these constraints should be in the form of viewpoint
tolerant (invariant and quasi-invariant) properties of the contours of the object
classes addressed. The properties themselves must characterize the projection of
objects at a hierarchy of features. This allows the design of rigorous segmentation
and grouping methods which proceed in a bottom-up fashion starting from the in-
tensity image and ending with high-level symbolic descriptions based on generalized
cylinder primitives and their relationships. The process is thus image-driven, but
guided by expectations derived from the generic representation scheme itself. The
challenges here are to effect robust and efficient segmentation and grouping from a
single intensity image in the presence of spurious information such as due to noise,
surface markings and shadows, and missing information such as due to occlusion
and feature fragmentation.
We then discuss the use of the extracted part-based descriptions to effect recogni-
tion from a database of object models. Such descriptions provide powerful indexing
and matching criteria. The challenges include the design of efficient strategies which
take full advantage of the (quasi) invariant, part-based and hierarchical, structure of
the descriptions and account for incomplete descriptions, shape dissimilarities and
uncertainty. Such strategies would result in recognition times nearly independent
of the size of the database and can be used as the basis for learning object models.
We discuss issues related to both the organization of the database and the recogni-
tion methods.
The discussion includes description of current work, with examples from recent
progress. We also discuss future research and the limitations inherent to the de-
scribed approach.
Pietro Perona
California Institute of Technology
and
University of Padova
perona@systems.caltech.edu
Abstract
We may classify an object as a car without knowing its make and model, as a face
without recognizing its owner. Objects belonging to the same class may look quite
different in detail; however, they often share some visual properties that allow us
to classify them based on their appearance.
I will present an approach to recognizing such visual classes. It is based on
combining in a probabilistic framework visual information about the appearance of
local features of an object with metric information about their spatial relationship.
Experimental results on an application of our ideas to detecting human faces in
cluttered scenes will be presented. Our algorithm is shown to have approximately
95% correct localization rate on a large, challenging database containing faces in
complicated and varied backgrounds.
In collaboration with M. Burl and T. Leung
Stanislava Simberova
Czech Academy of Sciences
Astrophysics Observatory Ondrejov
Czech Republic
ssimbero@asu.cas.cz
Abstract
Astronomy is an observational science. Unlike in a laboratory experiments, the
conditions cannot be changed. We can do nothing other than intercept the various
forms of energy. The study of electromagnetic radiation which includes visible light
is still the most dominant. The "new astronomy" subjects of radio, infrared and
X-ray have all developed rapidly in the last 20 years, stimulated in part by military
and other applications. Astronomical image processing applies variety of numerical
methods to extract scientific result from the observed data. The aim of the research
is to provide theoretical grounds and reliable methods for image processing in as-
tronomy and astrophysics. A major problem facing all of astronomy as a result of
the technological innovations is the data management. There are produced data,
mainly in the form of images (2-D) or spectra (1-D), in such voluminous quantities
as to make interactive analysis by human beings difficult. A great reliance must
be placed on computers. Already there is more international cooperation in the
development, support and sharing of astronomical software. Almost all astronomy,
even the "invisible" astronomy of X-rays and radio, depends on mapping in picto-
rial form for the human mind to visualize the pattern or distribution on the sky of
celestial sources emitting radiation at the wavelength of interest. This is done not
only to locate precisely the position of the source, but also to provide information
on its form or structure, and that of its environment, since much can be learned
from the appearance or morphology of certain objects. Images of the sky at differ-
ent wavelengths are essential for classification of objects, and accurate brightness
measurements at different wavelengths can also yield basic physical information.
Astronomical software for image processing
The special processing systems, coding standards and standard interfaces had been
developed for application programs in the last years. There are two image systems
widely used nowadays: MIDAS - Munich Image Data Analysis System, developed
for general processing and processing of the stellar image data; IDL - Interactive
Data Language, more oriented to the solar image data. Both systems involve the
current methods of image processing -
1. data reduction,
2. image enhancement,
3. image restoration,
4. image analysis methods.
Exchange of digital information between astronomical institutes led to the data
format standardization (FITS format).
New trends in astronomical image processing
Regarding to various type of data arising from the various sources, it is impossible
to give a general processing procedure. The methods and algorithms are developed
and discussed regarding to the goal of processing - an astrophysical interpretation.
At the Astronomical Institute of the Czech Academy of Sciences we are interested in
these areas: construction of the degradation model of the complex image informa-
tion system (telescope + image system) in general form; pattern recognition in the
images of solar photosphere - recognition in the internal structure of the solar spot
based on object moments and moment invariants; morphology of the photospheric
fine structure; image modeling using MRF. A new type of image reconstruction
method is developed very successfully. The image reconstruction is based on an
adaptive regression type of image destriping method using a regression model pre-
diction for replacing failed image row or isolated pixels. The results can be applied
also in other image acquisition and processing applications.
Shree K. Nayar and Hiroshi Murase
Department of Computer Science
Columbia University
New York, N.Y. 10027
nayar@cs.columbia.edu
Abstract
Vision research has laid significant emphasis on the development of compact and de-
scriptive shape representations for object recognition [Requicha 80 , Besl and Jain 85 ,
Nalwa 93 ]. This has lead to the creation of a variety of novel representations, in-
cluding, generalized cylinders [Binford 87 ], superquadrics [Barr 81 , Pentland 86 ],
extended gaussian images [Horn 84 ], parametric bicubic patches [Nalwa 93 ] and
differential geometric representations [Brady et al. 85 ], only to name a few. While
these representations are all useful in specific application domains, each has been
found to have its own drawbacks. This has kept researchers in search for more
powerful representations.
Will shape representation suffice? After all, vision deals with brightness images
that are functions not only of shape but also other intrinsic scene properties such as
reflectance and perpetually varying factors such as illumination. This observation
has motivated us to take an extreme approach to visual representation. What we
seek is not a representation of shape but rather appearance [Murase and Nayar 95 ],
encoded in which are brightness variations caused by three-dimensional shape, sur-
face reflectance properties, sensor parameters, and illumination conditions. Given
the number of factors at work, it is immediate that an appearance representation
that captures all possible variations is simply impractical. Fortunately, there exist
a wide collection of vision applications where pertinent variables are few and hence
compact appearance representation in a low-dimensional space is indeed possible.
An added drawback of shape representation emerges when a vision programmer
attempts to develop a practical recognition system. Techniques for automatically
acquiring shape models from sample objects are only being researched. For now,
a vision programmer is forced to select an appropriate shape representation, de-
sign object models using the chosen representation, and then manually input this
information into the system. This procedure is cumbersome and impractical when
dealing with large sets of objects, or objects with complex shapes. It is clear that
recognition systems of the future must be capable of acquiring object models with-
out human assistance. It turns out that the appearance representation proposed
here is easier to acquire through an automatic learning phase than to create man-
ually.
The appearance of an object is the combined effect of its shape, reflectance
properties, pose in the scene, and the illumination conditions. While shape and
reflectance are intrinsic properties that do not change for any rigid object, pose
and illumination vary from one scene to the next. We approach the visual learning
problem as one of acquiring a compact model of the object's appearance under
different poses and illumination directions. The object is "shown" to the image
sensor in several orientations and lighting conditions. This can be accomplished
using, for example, two robot manipulators; one rotates the object while the other
varies the illumination direction. The result is a large set of object images. These
images could either be used directly or after being processed to enhance object
characteristics. Since all images in the set are of the same object, consecutive
images are correlated to a large degree. The problem then is to compress this large
image set to a low-dimensional representation of object appearance.
A well-known image compression or coding technique is based on principal com-
ponent analysis, also known as the Karhunen-Loeve transform [Oja 83 ] [Fukunaga 90 ].
It uses the eigenvectors of an image set as orthogonal bases for representing indi-
vidual images in the set. Though a large number of eigenvectors may be required
for very accurate reconstruction of an object image, only a few are generally suffi-
cient to capture the significant appearance characteristics of an object, as shown in
[Sirovich and Kirby 87 , Turk and Pentland 91 ]. These eigenvectors constitute the
dimensions of what we refer to as the eigenspace. From the perspective of machine
vision, the eigenspace has an attractive property. If any two images from the set
are projected to the eigenspace, the distance between the corresponding points in
eigenspace is the best approximation to correlation between the images.
We have developed a continuous and compact representation of object appear-
ance that is parametrized by the variables, namely, object pose and illumination.
This representation is referred to as the parametric eigenspace [Murase and Nayar 93,
Murase and Nayar 95 ]. We have shown that parametric eigenspaces are useful not
only for object recognition but a variety of other vision tasks. For object recognition,
first an image set of the object is obtained by varying pose and illumination in small
increments. The image set is then normalized in brightness and scale to achieve in-
variance to sensor magnification and illumination intensity. The eigenspace for the
image set is constructed and all object images (learning samples) are projected to it
to obtain a set of points. These points lie on a manifold that is parametrized by pose
and illumination. The manifold is constructed from the discrete points by spline
interpolation [Murase and Nayar 95 ]. For the class of objects with linear reflectance
models, we have analyzed of effect of illumination on the structure of the manifold
[Nayar and Murase 94 ]. In was shown that, in the case of an ideal diffuse object
with arbitrary texture, three illumination directions are sufficient to construct the
entire illumination manifold. This result drastically reduces the number of images
required in the learning stage. It is important to note that the parametric appear-
ance representation of an object is acquired without prior knowledge of the object's
shape and reflectance properties. It is learned using just a sample of the object.
Recognition and pose estimation can be summarized as follows. Given an image
consisting of an object of interest, we assume that the object is not occluded and can
be segmented from the remaining scene. The segmented image region is normalized
in scale and brightness, such that it has the same size and brightness range as the
images used in the learning stage. This normalized image is projected to eigenspace.
The closest manifold reveals the identity of the object and exact position of the
closest point on the manifold determines pose and illumination direction. Two
different techniques have been tested for determining the closest manifold point,
one is based on binary search [Nene and Nayar 95 ] and other uses an input-output
mapping network [Mukherkee and Nayar 95 ]. We have achieved further speed-up in
recognition by developing a comprehensive theory and a novel algorithm for pattern
rejection [Baker and Nayar 95 ].
Will appearance representation suffice? Given the large number of parameters
that affect appearance, it does not suggest itself as a replacement for shape rep-
resentation. In fact, our experiments on recognition and robot tracking show that
appearance models are in many ways complementary to shape models. Appear-
ance representation proves extremely effective when the task variables are few; it
is efficient and circumvents time-consuming and often unreliable operations such
as feature detection. On the other hand, when occlusion effects are not negligi-
ble, shape models offer solutions in the form of partial matching that appearance
representation does not easily lend itself to.
Parametric appearance models have been applied to a variety of problems be-
sides object recognition, such as, illumination planning for robust recognition
[Murase and Nayar 94a ] [Murase and Nayar 94b ], visual positioning and tracking
[Nayar et al. 94 ], and temporal inspection of complex parts [Nayar et al. 95 ]. These
applications have demonstrated that the techniques underlying appearance model-
ing and matching are general. This motivated us to develop a comprehensive soft-
ware package [Nene et al. 94 ] for appearance matching that is presently being used
at several research institutions.
References
[Baker and Nayar 95] S. Baker and S. K. Nayar, "A Theory of Pattern Rejection,"
Tech. Rep. CUCS-013-95, Dept. of Computer Science, Columbia Univ., May
1995.
[Barr 81] A. H. Barr, "Superquadric and Angle Preserving Transformations," IEEE
Computer Graphics and Applications, Vol. 1, No. 1, pp. 11-23, Jan. 1981.
[Besl and Jain 85] P. J. Besl and R. C. Jain, "Three-Dimensional Object Recogni-
tion," Computing Surveys, Vol. 17, No. 1, pp. 75-145, Mar. 1985.
[Binford 87] T. O. Binford, "Generalized Cylinder Representation," Encyclopedia
of Artificial Intelligence, S. C. Shapiro, Ed., John Wiley & Sons, New York,
pp. 321-323, 1987.
[Brady et al. 85] M. Brady, J. Ponce, A. Yuille and H. Asada, "Describing Sur-
faces," Computer Vision, Graphics, and Image Processing, Vol. 32, pp. 1-28,
1985.
[Chin 86] R. T. Chin and C. R. Dyer, "Model-Based Recognition in Robot Vision,"
ACM Computing Surveys, Vol. 18, No. 1, pp. 67-108, 1986.
[Fukunaga 90] K. Fukunaga, Introduction to Statistical Pattern Recognition, Aca-
demic Press, London, 1990.
[Horn 84] B. K. P. Horn, "Extended Gaussian Images," Proceedings of the IEEE,
Vol. 72, No. 12, pp. 1671-1686, Dec. 1984.
[Mukherkee and Nayar 95] S. Mukherjee and S. K. Nayar, "Optimal RBF Networks
for Visual Learning," Proc. of International Conference on Computer Vision,
Boston, June 1995.
[Murase and Nayar 93] H. Murase and S. K. Nayar, "Learning Object Models from
Appearance," Proc. of AAAI, Washington D. C., July 1993.
[Murase and Nayar 94a] H. Murase and S. K. Nayar, "Illumination Planning for
Object Recognition in Structured Environments," Proc. of IEEE Intl. Conf.
on Robotics and and Automation, Seattle, pp. 31-38, June 1994.
[Murase and Nayar 94b] H. Murase and S. K. Nayar, "Illumination Planning for
Object Recognition Using Parametric Eigenspaces," IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 16, No. 12, pp. 1219-1227
Dec., 1994.
[Murase and Nayar 95] H. Murase and S. K. Nayar, "Visual Learning and Recog-
nition of 3D Objects from Appearance," International Journal of Computer
Vision, Vol. 14, No. 1, pp. 5-24, 1995.
[Nalwa 93] V. S. Nalwa, A Guided Tour of Computer Vision, Addison Wesley, 1993.
[Nayar et al. 94] S. K. Nayar, H. Murase and S. A. Nene, "Learning, Positioning,
and Tracking Visual Appearance," Proc. of IEEE Intl. Conf. on Robotics and
Automation, San Diego, May 1994.
[Nayar and Murase 94] S. K. Nayar and H. Murase, "On the Dimensionality of Illu-
mination in Eigenspace," Tech. Rep. CUCS-021-94, Dept. of Computer Science,
Columbia Univ., Aug. 1994.
[Nayar et al. 95] S. K. Nayar, S. A. Nene and H. Murase, "Subspace Methods for
Robot Vision," Tech. Rep. CUCS-06-95, Dept. of Computer Science, Columbia
Univ., Mar. 1995.
[Nene et al. 94] S. A. Nene, S. K. Nayar and H. Murase, "SLAM: A Software Li-
brary for Appearance Matching," Proc. of ARPA IU Workshop, Monterey,
Nov. 1994.
[Nene and Nayar 95] S. A. Nene and S. K. Nayar, "Binary Search Through Multiple
Dimensions," Tech. Rep. CUCS-018-94, Dept. of Computer Science, Columbia
Univ., June 1994, Revised August 1995.
[Oja 83] E. Oja, Subspace methods of Pattern Recognition, Res. Studies Press, Hert-
fordshire, 1983.
[Pentland 86] A. P. Pentland, "Perceptual Organization and the Representation of
Natural Form,' ' Artificial Intelligence, Vol. 28, pp. 293-331, 1986.
[Requicha 80] A. A. G. Requicha, "Representation of Rigid Solids: Theory, Meth-
ods and Systems," Computing Surveys, Vol. 12, No. 4, pp. 1-437-464, December
1980.
[Sirovich and Kirby 87] L. Sirovich and M. Kirby, "Low dimensional procedure for
the characterization of human faces," Journal of Optical Society of America,
Vol. 4, No. 3, pp. 519-524, 1987.
[Turk and Pentland 91] M. A. Turk and A. P. Pentland, "Face Recognition Using
Eigenfaces," Proc. of IEEE CVPR, pp. 586-591, June 1991.
Vaclav Hlavac and Tomas Werner
Computer Vision Laboratory
Czech Technical University
Karlovo nam. 13
CZ 121 35 Praha 2
hlavac@vision.felk.cvut.cz
Abstract
The talk will present a new approach to rendering arbitrary views of real-world 3-D
objects of complex shapes, captured by an ordinary TV camera. We propose to
represent an object by a small set of corresponding 2-D views, and to construct
any other view as a combination of these views. We show that this combination
can be linear, assuming proximity of the views, and we suggest how the visibility of
constructed points can be solved. Our approach entirely eliminates the need for the
difficult 3-D model reconstruction. We present results on real objects, indicating
that our approach is feasible.
Josef Honec and Pavel Zemcik
Technical University Brno
Dept. of Automation and Measurement
Bozetechova 2
Brno, Czech Republic
honec@dame.fee.vutbr.cz
Abstract
The contribution discusses features and application limitations of current CCD
video cameras from the view point of the resolution, inaccuracies of the optical
systems, video signal processing, and features of the digitizers.
Efficient methods for measuring and dealing with the inaccuracies of the camera
and lens systems.
The achieved result are illustrated on practical examples of systems that exploit
dimensions measurement, 3D reconstruction, and monitoring of traffic situations.
Some of the systems are already being used in the industry.
Antony Hoogs and Ruzena Bajcsy
Will be added soon!
Narendra Ahuja
Will be added soon!
Takeo Kanade
Will be added soon!
S. Lin and Sang Wong Lee
Will be added soon!