Single View Metrology

A. Criminisi, I. Reid and A. Zisserman 

Department of Engineering Science
University of Oxford
Oxford, UK, OX1 3PJ
fcriminis,ian,[email protected]

Abstract direction (not parallel to the plane). We are then concerned

with three canonical types of measurement: (i) measure-
We describe how 3D affine measurements may be com- ments of the distance between any of the planes which are
puted from a single perspective view of a scene given only parallel to the reference plane; (ii) measurements on these
minimal geometric information determined from the image. planes (and comparison of these measurements to those ob-
This minimal information is typically the vanishing line of tained on any plane); and (iii) determining the camera’s po-
a reference plane, and a vanishing point for a direction not sition in terms of the reference plane and direction. The
parallel to the plane. It is shown that affine scene structure measurement methods developed here are independent of
may then be determined from the image, without knowledge the camera’s internal parameters: focal length, aspect ratio,
of the camera’s internal calibration (e.g. focal length), nor principal point, skew.
of the explicit relation between camera and world (pose). The ideas in this paper can be seen as reversing the rules
In particular, we show how to (i) compute the distance for drawing perspective images given by Leon Battista Al-
between planes parallel to the reference plane (up to a com- berti [1] in his treatise on perspective (1435). These are
mon scale factor); (ii) compute area and length ratios on the rules followed by the Italian Renaissance painters of the
any plane parallel to the reference plane; (iii) determine the 15th century, and indeed we demonstrate the correctness
camera’s (viewer’s) location. Simple geometric derivations of their mastery of perspective by analysing a painting by
are given for these results. We also develop an algebraic Piero della Francesca.
representation which unifies the three types of measurement We begin in section 2 by giving geometric interpretations
and, amongst other advantages, permits a first order error for the key scene features, and then give simple geomet-
propagation analysis to be performed, associating an un- ric derivations of how, in principle, three dimensional affine
certainty with each measurement. information may be extracted from the image. In section
We demonstrate the technique for a variety of applica- 3 we introduce an algebraic representation of the problem
tions, including height measurements in forensic images and show that this representation unifies the three canoni-
and 3D graphical modelling from single images. cal measurement types, leading to simple formulae in each
case. In section 4 we describe how errors in image mea-
surements propagate to errors in the 3D measurements, and
1. Introduction
hence we are able to compute confidence intervals on the 3D
In this paper we describe how aspects of the affine 3D measurements, i.e. a quantitative assessment of accuracy.
geometry of a scene may be measured from a single per- The work has a variety of applications, and we demonstrate
spective image. We will concentrate on scenes containing two important ones: forensic measurement and virtual mod-
planes and parallel lines, although the methods are not so elling in section 5.
restricted. The methods we develop extend and generalize
previous results on single view metrology [8, 9, 13, 14]. 2. Geometry
It is assumed that images are obtained by perspective
projection. In addition, we assume that the vanishing line of The camera model employed here is central projection.
a reference plane in the scene may be determined from the We assume that the vanishing line of a reference plane in
image, together with a vanishing point for another reference the scene may be computed from image measurements, to-
 The authors would like to thank Andrew Fitzgibbon for assistance gether with a vanishing point for another direction (not par-
with the TargetJr libraries, and David Liebowitz and Luc van Gool for dis- allel to the plane). This information is generally easily ob-
cussions. This work was supported by the EU Esprit Project IMPROOFS. tainable from images of structured scenes [3, 11, 12]. Ef-
 plane vanishing line camera centre v
vanishing point
l
ref.
dir.
image plane vanishing
point
v
i
l vanishing line

reference plane
t
π /

l
Figure 1: Basic geometry: The plane’s vanishing line is the intersection b
of the image plane with a plane parallel to the reference plane and passing π
v
through the camera centre. The vanishing point is the intersection of
the image plane with a line parallel to the reference direction through the
camera centre.

b
Figure 2: Cross ratio: The point on the plane  corresponds to the
fects such as radial distortion (often arising in slightly wide- t v
point on the plane  0 . They are aligned with the vanishing point . The
angle lenses typically used in security cameras) which cor- vtb i
four points , , and the intersection of the line joining them with the
rupt the central projection model can generally be removed vanishing line define a cross-ratio. The value of the cross-ratio determines
a ratio of distances between planes in the world, see text.
[6], and are therefore not detrimental to our methods (see,
for example, figure 9).
Although the schematic figures show the camera centre of the cross-ratio provides an affine length ratio. In fact we
at a finite location, the results we derive apply also to the obtain the ratio of the distance between the planes contain-
case of a camera centre at infinity, i.e. where the images are t b
ing and , to the camera’s distance from the plane  (or
obtained by parallel projection. The basic geometry of the 0 depending on the ordering of the cross-ratio). The abso-
plane’s vanishing line and the vanishing point are illustrated lute distance can be obtained from this distance ratio once
l
in figure 1. The vanishing line of the reference plane is the the camera’s distance from  is specified. However it is
projection of the line at infinity of the reference plane into usually more practical to determine the distance via a sec-
v
the image. The vanishing point is the image of the point ond measurement in the image, that of a known reference
at infinity in the reference direction. Note that the reference length.
direction need not be vertical, although for clarity we will Furthermore, since the vanishing line is the imaged axis
often refer to the vanishing point as the “vertical” vanishing of the pencil of planes parallel to the reference plane, the
point. The vanishing point is then the image of the vertical knowledge of the distance between any pair of the planes
“footprint” of the camera centre on the reference plane. is sufficient to determine the absolute distance between an-
It can be seen (for example, by inspection of figure 1) other two of the planes.
that the vanishing line partitions all points in scene space. Example. Figure 3 shows that a person’s height may be
Any scene point which projects onto the vanishing line is computed from an image given a vertical reference height
at the same distance from the plane as the camera centre; elsewhere in the scene. The formula used to compute this
if it lies “above” the line it is further from the plane, and if result is given in section 3.1.
“below” the vanishing line, then it is closer to the plane than
the camera centre. 2.2. Measurements on parallel planes
Two points on separate planes (parallel to the reference
plane) correspond if the line joining them is parallel to the If the reference plane  is affine calibrated (we know
reference direction; hence the image of each point and the its vanishing line) then from image measurements we can
vanishing point are collinear. For example, if the direction compute: (i) ratios of lengths of parallel line segments on
is vertical, then the top of an upright person’s head and the the plane; (ii) ratios of areas on the plane. Moreover the
sole of his/her foot correspond. vanishing line is shared by the pencil of planes parallel to
the reference plane, hence affine measurements may be ob-
2.1. Measurements between parallel planes tained for any other plane in the pencil. However, although
affine measurements, such as an area ratio, may be made on
We wish to measure the distance between two parallel a particular plane, the areas of regions lying on two parallel
t
planes, specified by the image points and , in the refer- b planes cannot be compared directly. If the region is parallel
ence direction. Figure 2 shows the geometry, with points projected in the scene from one plane onto the other, affine
t b
and in correspondence. The four points marked on the measurements can then be made from the image since both
figure define a cross-ratio. The vanishing point is the image regions are now on the same plane, and parallel projection
of a point at infinity in the scene [15]. In the image the value between parallel planes does not alter affine properties.
 v

t
/ π /
l
X x
π
/
/
T
b
x
π π
X
B

Figure 4: Homology mapping between parallel planes: (left) A point

X
X on plane  is mapped into the point 0 on  0 by a parallel projection.
(right) In the image the mapping between the images of the two planes is
v l
a homology, with the vertex and the axis. The correspondence ! b t
fixes the remaining degree of freedom of the homology from the cross-ratio
vit
of the four points: , , and . b
t tr
i ir
178.8 cm

b br

Figure 3: Measuring the height of a person: (top) original image; (bot-

tom) the height of the person is computed from the image as 178.8cm (the
true height is 180cm, but note that the person is leaning down a bit on his
right foot). The vanishing line is shown in white and the reference height
t b
is the segment ( r ; r ). The vertical vanishing point is not shown since it
t b
lies well below the image. is the top of the head and is the base of the
i
feet of the person while is the intersection with the vanishing line.

A map in the world between parallel planes induces a

map in the image between images of points on the two
planes. This image map is a planar homology [15], which is
a plane projective transformation with five degrees of free-
dom, having a line of fixed points, called the axis and a Figure 5: Measuring the ratio of lengths of parallel line segments lying
distinct fixed point not on the axis known as the vertex. Pla- t b
on two parallel scene planes: The points and (together with the plane
nar homologies arise naturally in an image when two planes vanishing line and the vanishing point) define the homology between the
two planes on the facade of the building.
related by a perspectivity in 3-space are imaged [16]. The
geometry is illustrated in figure 4. ogy and then, since the reference plane’s vanishing line is
In our case the vanishing line of the plane, and the verti- known, make affine measurements in the plane, e.g. parallel
cal vanishing point, are, respectively, the axis and vertex of length or area ratios.
the homology which relates a pair of planes in the pencil. Example. Figure 5 shows that given the reference vanish-
This line and point specify four of the five degrees of free- ing line and vanishing point, and a point correspondence (in
dom of the homology. The remaining degree of freedom of the reference direction) on each of two parallel planes, then
the homology is uniquely determined from any pair of im-
age points which correspond between the planes (points b the ratio of lengths of parallel line segments may be com-
t
and in figure 4).
puted from the image. The formula used to compute this
result is given in section 3.2.
This means that we can compare measurements made
on two separate planes by mapping between the planes in
2.3. Determining the camera position
the reference direction via the homology. In particular we
may compute (i) the ratio between two parallel lengths, one In section 2.1, we computed distances between planes as
length on each plane; (ii) the ratio between two areas, one a ratio relative to the camera’s distance from the reference
area on each plane. In fact we can simply transfer all points plane. Conversely, we may compute the camera’s distance
from one plane to the reference plane using the homol- from a particular plane knowing a single reference distance.
 Furthermore, by considering figure 1 it is seen that the plane to image map is degenerate. Consequently, the final
location of the camera relative to the reference plane is the column (the origin of the coordinate system) must not lie
back-projection of the vanishing point onto the reference on the vanishing line, since if it does then all three columns
plane. This back-projection is accomplished by a homog- are points on the vanishing line, and thus are not linearly
raphy which maps the image to the reference plane (and o p l l
independent. Hence we set it to be = 4 = =jj jj = . ^l
vice-versa). Although the choice of coordinate frame in the Therefore the final parametrization of the projection ma-
world is somewhat arbitrary, fixing this frame immediately trix P is:
v ^l
 
defines the homography uniquely and hence the camera po- P = l?
1 l?2 (1)
sition.
We show an example in figure 12, where the location of where is a scale factor, which has an important rôle to
the camera centre has been determined, and superimposed play in the remainder of the paper.
into a virtual view of the scene. In the following sections we show how to compute
various measurements from this projection matrix. Mea-
3. Algebraic Representation surements between planes are independent of the first two
(under-determined) columns of P. For these measurements
The measurements described in the previous section are the only unknown quantity is . Coordinate measurements
computed in terms of cross-ratios. In this section we de- within the planes depend on the first two and the fourth
velop a uniform algebraic approach to the problem which columns of P. They define an affine coordinate frame within
has a number of advantages over direct geometric construc- the plane. Affine measurements (e.g. area ratios), though,
tion: first, it avoids potential problems with ordering for the are independent of the actual coordinate frame and depend
cross-ratio; second, it enables us to deal with both mini- only on the fourth column of P. If any metric information
mal or over-constrained configurations uniformly; third, we on the plane is known, we may impose constraints on the
unify the different types of measurement within one rep- choice of the frame.
resentation; and fourth, in section 4 we use this algebraic
representation to develop an uncertainty analysis for mea- 3.1. Measurements between parallel planes
surements.
To begin we define an affine coordinate system XY Z in We wish to measure the distance between scene planes
space. Let the origin of the coordinate frame lie on the refer- specified by a base point B on the reference plane and top
ence plane, with the X and Y -axes spanning the plane. The point T in the scene. These points may be chosen as respec-
Z -axis is the reference direction, which is thus any direc- tively (X; Y; 0) and (X; Y; Z ), and their images are andb
tion not parallel to the plane. The image coordinate system t . If P is the projection matrix then the image coordinates
is the usual xy affine image frame, and a point X in space is are 2 3 2 3
x
projected to the image point via a 3  4 projection matrix X X
6Y 7 6Y 7
P as:   b= 4 5;
P6 7
0 t= P6 7
4Z 5
x = PX = p1 p2 p3 p4 X
1 1
where x and X are homogeneous vectors in the form:
x = (x; y; w)> , X = (X; Y; Z; W )> , and ‘=’ means The equations above can be rewritten as
equality up to scale. b = (X p1 + Y p2 + p4 ) (2)
If we denote the vanishing points for the X , Y and Z
v v v
directions as (respectively) X , Y and , then it is clear by
t = (X p1 + Y p2 + Z p3 + p4) (3)
inspectionthat the first three columns of P are the vanishing where  and  are unknown scale factors, and pi is the ith
v p v
points; X = 1 , Y = 2 and p v p = 3 , and that the column of the P matrix.
final column of P is the projection of the origin of the world Taking the scalar product of (2) with ^l yields  = ^l
b,
o p
coordinate system, = 4 . Since our choice of coordinate and combining this with the third column of (1) and (3) we
p
frame has the X and Y axes in the reference plane 1 = X v obtain
p v
and 2 = Y are two distinct points on the vanishing line. ,jjb  tjj
Choosing these points fixes the X and Y affine coordinate Z= ^
(l
b)jjv  tjj
(4)
l
axes. We denote the vanishing line by , and to emphasise
v v
that the vanishing points X and Y lie on it, we denote Since Z scales linearly with we have obtained affine
them by ?l l ? ?
l l
1 , 2 , with i
= 0. structure. If is known, then we immediately obtain a met-
Columns 1, 2 and 4 of the projection matrix are the three ric value for Z . Conversely, if Z is known (i.e. it is a refer-
columns of the reference plane to image homography. This ence distance) then we have a means of computing , and
homography must have rank three, otherwise the reference hence removing the affine ambiguity.
Figure 6: Measuring heights using parallel lines: Given the vertical
vanishing point, the vanishing line for the ground plane and a reference
height, the distance of the top of the window on the right wall from the
ground plane is measured from the distance between the two horizontal
Figure 7: Measuring ratios of areas on separate planes: The image
lines shown, one defined by the top edge of the window, and the other on
the ground plane. t b
points and together with the vanishing line of the two parallel planes
and the vanishing point for the orthogonal direction define the homology
between the planes. The ratio between the area of the window on the left
Example. In figure 6 heights from the ground plane are plane and that of the window on the right plane is computed.
measured between two parallel lines, one off the plane (top)
and one on the plane (base). In fact, thanks to the plane 3.3. Determining camera position
vanishing line, given one line parallel to the reference plane
it is easy to compute the family of parallel lines. Computing Suppose the camera centre is C = (Xc ; Yc ; Zc ; Wc )> in
the distance between them is a straightforward application affine coordinates (see figure 1). Then since PC = 0 we
of (4). have

3.2. Measurements on parallel planes

PC = l? ?
1 Xc + l2 Yc + vZc + ^lWc = 0 (6)

The projection matrix P from the world to the image is The solution to this set of equations is given (using Cramer’s
defined above with respect to a coordinate frame on the ref- rule) by
erence plane. In this section we determine the projection 
Xc = ,det l?2 v ^l, Yc = det l?1 v ^l 
matrix P0 referred to the parallel plane  0 and we show how ^l , Wc = det l?1 l?2 v
Zc = ,det l?1 l?2 (7)
the homology between the two planes can be derived di-
rectly from the two projection matrices. Note that once again we obtain structure off the plane up to
Suppose the world coordinate system is translated from
the plane  onto the plane 0 along the reference direction,
the affine scale factor . As before, we may upgrade the
distance to metric with knowledge of , or use knowledge
then it is easy to show that we can parametrize the new pro-
jection matrix P0 as:
of camera height to compute and upgrade the affine struc-
ture.
Z v + ^l
 
P0 = p1 p2 v Note that affine viewing conditions (where the camera
centre is at infinity) present no problem to the expressions
>
where Z is the distance between the planes. Note that if in (7), since in this case we have = 0 0  ^l

Z = 0 then P0 = P correctly.
and
 >
The plane to image homographies can be extracted from v=   0 . Hence Wc = 0 so we obtain a cam-
the projection matrices ignoring the third column, to give: era centre on the plane at infinity, as we would expect. This
point on  1 represents the viewing direction for the paral-
p1 p2 ^l ; Z v + ^l
   
H= H0 = p1 p2 lel projection.
Then H~ = H0 H,1 maps image points on the plane  onto If the viewpoint is finite (i.e. not affine viewing condi-
tions) then the formula for Zc may be developed further
points on the plane  0 and so defines the homology.
~ as: by taking the scalar product of both sides of (6) with the
^l ^l v
vanishing line . The result is: Zc = ,(
),1 .
A short computation gives the homology matrix H

H~ = I + Z vl^> (5)
4. Uncertainty Analysis
Given the homology between two planes in the pencil we
can transfer all points from one plane to the other and make Feature detection and extraction – whether manual or au-
affine measurements in the plane (see fig 5 and fig 7). tomatic (e.g. using an edge detector) – can only be achieved
 t Λt ^t

Figure 9: Uncertainty analysis on height measurements: The image

shown was captured from a cheap security type camera which exhibited
radial distortion. This has been corrected and the height of the man es-

b Λb ^b timated (measurements are in cm). (left) The height of the man and the
associated uncertainty are computed as 190.6cm (c.f. ground truth value
190cm). The vanishing line for the ground plane is shown in white at
the top of the image. When one reference height is used the uncertainty
Figure 8: Maximum likelihood estimation of the top and base points (3-sigma) is 4:1cm, while (right) it reduces to 2:9cm as two more ref-
(closeup of fig. 9): (left) The top and base uncertainty ellipses, respec- erence heights are introduced (the filing cabinet and the table on the left).
tively t and b , are shown. These ellipses are specified by the user, and
indicate a confidence region for localizing the points. (right) MLE top and Now, assuming the statistical independence of and P ^z
t b
base points ^ and ^ are aligned with the vertical vanishing point (outside we obtain a first order approximation for the variance of the
the image).
distance measurement:
 
to a finite accuracy. Any features extracted from an image, 2 
h = rh 0  rh >^z 0
(8)
therefore, are subject to errors. In this section we consider P
how these errors propagate through the measurement for-
mulae in order to quantify the uncertainty on the final mea- where rh is the 1  10 Jacobian matrix of the function
surements. which maps the projection matrix and top and base points to
When making measurements between planes, uncer- a distance between them (4). The validity of all approxima-
tainty arises from the uncertainty in P, and from the uncer- tion has been tested by Monte Carlo simulations and by a
tain image locations of the top and base points and . The t b number of measurements on real images where the ground
uncertainty in P depends on the location of the vanishing truth was known.
line, the location of the vanishing point, and on , the affine Example. An image obtained from a poor quality security
scale factor. Since only the final two columns contribute, camera is shown in figure 9. It has been corrected for ra-
we model the uncertainty in P as a 6  6 homogeneous co- dial distortion using the method described in [6], and the
variance matrix, P . Since the two columns have only five floor taken as the reference plane. Vertical and horizontal
v
degrees of freedom (two for , two for and one for ),l lines are used to compute the P matrix of the scene. One
the covariance matrix is singular, with rank five. Details reference height is used to obtain the affine scale factor
of its computation are given in [4] and are omitted here for from (4), so other measurements in the same direction are
brevity. metric.
Likewise, the uncertainty in the top and base points (re- The computed height of the man and an associated 3-
sulting largely from the finite accuracy with which these standard deviation uncertainty are displayed in the figure.
features may be located in the image) is modelled by covari- The height obtained differs by only 6mm from the known
ance matrices b and t . Since in the error-free case, these true value. As the number of reference distances is in-
points must be aligned with the vertical vanishing point we creased, so the uncertainty on P (in fact just on ) de-
can determine maximum likelihood estimates of their true creases, resulting in a decrease in uncertainty of the mea-
locations (^ and ^ ) by minimising the objective
t b sured height, as theoretically expected.

(b2 , b^2 )> ,b21 (b2 , b^2 ) + (t2 , ^t2 )> ,t21 (t2 , ^t2 )
5. Applications
(which is the sum of the Mahalanobis distances between
the input points and the ML estimates, the subscript 2 in-
5.1. Forensic science
dicates inhomogeneous 2-vectors) subject to the alignment
v ^t b^
A common requirement in surveillance images is to ob-
constraint
(  ) = 0. Using standard techniques [7] tain measurements from the scene, such as the height of a
we obtain a first order approximation to the 4  4, rank three felon. Although, the felon has usually departed the scene,
covariance of the parameters = ( ^> ^z2
^ >2 )> . Figure 8
t b reference lengths can be measured from fixtures such as ta-
illustrates the idea. bles and windows.
Figure 10: Measuring the height of a person in an outdoor scene: The
ground plane is the reference plane, and its vanishing line is computed
from the slabs on the floor. The vertical vanishing point is computed from
the edges of the phone box, whose height is known and used as reference.
The veridical height is 187cm, but note that the person is leaning slightly
on his right foot.

Figure 12: Complete 3D reconstruction of a real scene: (left) original

Figure 11: Measuring heights of objects on separate planes: Using the image; (right) a view of the reconstructed 3D model; (bottom) A view of
homology between the ground plane (initial reference) and the plane of the the reconstructed 3D model which shows the position of the camera centre
table, we can determine the height of the file on the table. (plane location X,Y and height) with respect to the scene.

In figure 10 the edges of the paving stones on the floor lengths. The position of the camera centre is also estimated
are used to compute the vanishing line of the ground plane; and is shown in the figure.
the edges of the phonebox to compute the vertical vanishing
point; and the height of the phone box provides the metric 5.3. Modelling paintings
calibration in the vertical direction. The height of the person
is then computed using (4). Figure 13 shows a masterpiece of Italian Renaissance
Figure 11 shows an example where the homology is used painting, “La Flagellazione di Cristo” by Piero della
to project points between planes so that a vertical distance Francesca (1416 - 1492). The painting faithfully follows the
may be measured given the distance between a plane and geometric rules of perspective, and therefore we can apply
the reference plane. the methods developed here to obtain a correct 3D recon-
struction of the scene.
5.2. Virtual modelling Unlike other techniques [8] whose main aim is to cre-
ate convincing new views of the painting regardless of the
In figure 12 we show an example of complete 3D re- correctness of the 3D geometry, here we reconstruct a geo-
construction of a scene. Two sets of horizontal edges are metrically correct 3D model of the viewed scene.
used to compute the vanishing line for the ground plane, In the painting analyzed here, the ground plane is chosen
and vertical edges used to compute the vertical vanishing as reference and its vanishing line can be computed from
point. Four points with known Euclidean coordinates deter- the several parallel lines on it. The vertical vanishing point
mine the metric calibration of the ground plane and thus for follows from the vertical lines and consequently the relative
the pencil of horizontal planes which share the vanishing heights of people and columns can be computed. Further-
line. The distance of the top of the window to the ground, more the ground plane can be rectified from the square floor
and the height of one of the pillars are used as reference patterns and therefore the position on the ground of each
vertical object estimated [5, 10]. The measurements, up to
an overall scale factor, are used to compute a three dimen-
sional VRML model of the scene. Two different views of
the model are shown in figure 13.

6. Summary and Conclusions

We have explored how the affine structure of 3-space
may be partially recovered from perspective images in
terms of a set of planes parallel to a reference plane and a
reference direction not parallel to the reference plane. More
generally, affine 3 space may be represented entirely by sets
of parallel planes and directions [2]. We are currently in-
vestigating how this full geometry is best represented and
computed from a single perspective image.

References
[1] L. B. Alberti. De Pictura. Laterza, 1980.
[2] M. Berger. Geometry II. Springer-Verlag, 1987.
[3] R. T. Collins and R. S. Weiss. Vanishing point calculation
as a statistical inference on the unit sphere. In Proc. ICCV,
pages 400–403, Dec 1990.
[4] A. Criminisi, I. Reid, and A. Zisserman. Computing 3D eu-
clidean distance from a single view. Technical Report OUEL
2158/98, Dept. Eng. Science, University of Oxford, 1998.
[5] A. Criminisi, I. Reid, and A. Zisserman. A plane measuring
device. Image and Vision Computing, 17(8):625–634, 1999.
[6] F. Devernay and O. Faugeras. Automatic calibration and re-
moval of distortion from scenes of structured environments.
In SPIE, volume 2567, San Diego, CA, Jul 1995.
[7] O. Faugeras. Three-Dimensional Computer Vision: a Geo-
metric Viewpoint. MIT Press, 1993.
[8] Y. Horry, K. Anjyo, and K. Arai. Tour into the picture: Using
a spidery mesh interface to make animation from a single
image. In Proc. ACM SIGGRAPH, pages 225–232, 1997.
[9] T. Kim, Y. Seo, and K. Hong. Physics-based 3D position
analysis of a soccer ball from monocular image sequences.
Proc. ICCV, pages 721 – 726, 1998.
[10] D. Liebowitz, A. Criminisi, and A. Zisserman. Creating ar-
chitectural models from images. In Proc. EuroGraphics, Sep
1999.
[11] D. Liebowitz and A. Zisserman. Metric rectification for per-
spective images of planes. In Proc. CVPR, pages 482–488,
Jun 1998.
[12] G. F. McLean and D. Kotturi. Vanishing point detection by
line clustering. IEEE T-PAMI, 17(11):1090–1095, 1995.
[13] M. Proesmans, T. Tuytelaars, and L. Van Gool. Monoc- Figure 13: Complete 3D reconstruction of a Renaissance painting:
ular image measurements. Technical Report Improofs- (top) La Flagellazione di Cristo, (1460, Urbino, Galleria Nazionale delle
M12T21/1/P, K.U.Leuven, 1998. Marche). (middle) A view of the reconstructed 3D model. The patterned
[14] I. Reid and A. Zisserman. Goal-directed video metrology. In floor has been reconstructed in areas where it is occluded by taking advan-
R. Cipolla and B. Buxton, editors, Proc. ECCV, volume II, tage of the symmetry of its pattern. (bottom) another view of the model
pages 647–658. Springer, Apr 1996. with the roof removed to show the relative positions of people and archi-
[15] C. E. Springer. Geometry and Analysis of Projective Spaces. tectural elements in the scene. Note the repeated geometric pattern on the
floor in the area delimited by the columns (barely visible in the painting).
Freeman, 1964.
Note that the people are represented simply as flat silhouettes since it is not
[16] L. Van Gool, M. Proesmans, and A. Zisserman. Planar ho- possible to recover their volume from one image, they have been cut out
mologies as a basis for grouping and recognition. Image and manually from the original image. The columns have been approximated
Vision Computing, 16:21–26, Jan 1998. with cylinders.