HAL Id: inria-00598842
https://hal.inria.fr/inria-00598842
Submitted on 7 Jun 2011
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
Reconstruction de scènes à partir d’une seule image
Marta Wilczkowiak
To cite this version:
Marta Wilczkowiak. Reconstruction de scènes à partir d’une seule image. Digital Libraries [cs.DL].
2000. �inria-00598842�
U.F.R.
Informatique &MathématiquesAppliquées
WydzialMatematyki
iNaukInformaty zny h
DEA IMAGERIE VISION ROBOTIQUE
Proje t presented by
Marta Wil zkowiak
3D model a quisition from single image
Prepared in the INRIA Rhne-Alpes
in the MOVI proje t
Date: JUNE 2000
Supervisors: Dr Edmond Boyer
Prof. Krzysztof Mar iniak
1 Introdu tion 3
1.1 State of the art . . . 4
1.2 Our approa h . . . 6
1.3 Overview of this report . . . 7
Introdu tion
The human eye, onne ted with the human mind, is the base tool that
allows us to per eive and interpret the world surrounding us. Our eyes let
us not only simply see obje ts we an re ognize forms, followmovements,
estimate distan es and velo ities. The omputer vision is the dis ipline,
whi h triesto apply those apabilitiesto omputers. This allows toimprove
and omplete our vision and, on the other hand, to eliminate the need of
human intera tion in ertain appli ations.
The number and importan e of appli ations of omputer vision in
to-day's world is di ult to underestimate. Surveillan e, roboti s, assistan e
in surgi al operations and CAD/CAM systems are only some in the wide
range of dis iplines whi h are developed mostly thanks to progress in
om-puter vision. The reasons for applying omputer vision are numerous: its
exibility, ability of fast treatment of large amounts of data and, above all,
the possibilityto express informationina lear and intuitive way.
One bran h of omputer visiondeals with the extra tion of
threedimen-sional information from images. The ability of interpretation of 3D data is
ne essary insome kinds of appli ations. Consider for example roboti s, the
problem of obsta leavoidan e, or appli ations for3D modelvisualisation.
Inthisworkweareinterestedinusing omputervisionto onstru t
three-dimensional models of s enes. Re onstru tion of geometri models allows us
to observe s enes from dierent points of view. Texture information, whi h
an be extra ted from images, permits to obtain photorealisti alrenderings
approa h is are onstru tion of the 3Dmodelwith the use of several images
of thes ene. Butweare notalways abletoobtaintwoormoreimagesofthe
s ene. Let us onsider for example images oming from paintings, ar hives
et .
On the otherhand some spe ial, additionalinformationabout the s ene
anprovidestrong onstraintsonthe ameraparametersandpositionsofthe
model points in the 3D spa e. Espe ially useful are geometri al
dependen- ies, like parallelism,perpendi ularity, in iden e of lines orpointsand lines.
S enes, whi h provide data of this type, are for example urban or indoor
s enes.
Theaimofthisproje tistore onstru tthe3Dmodelofthes enehaving
in disposition only one image. Ne essary steps to doso are amera
alibra-tion and extra tion of the 3D textured information. We will be interested
mainly by images of indoor and outdoor ar hite tural s enes. It is easy to
observe, that su h s enes usually ontain some ommon elements, like
dif-ferent types of ubes, triangle prisms, solids built of sets of parallelograms
et . Inthis workwewillstudythe dierentamountsofinformationprovided
by proje tions of su h primitivesin the image and exploit them for amera
alibration and re onstru tion of the 3D modelof the s ene.
1.1 State of the art
The problems of amera alibration and 3D re onstru tion have been
studied for a long time. Proposed approa hes make use, in general, of two
or moreimages,either alibratedor not,whi hlead toeither Eu lideanor a
proje tive re onstru tion. They are explained forexample inFaugerasbook
[Fau93℄.
During the last years few approa hes to the problem of Eu lidean
re- onstru tion from a single image where introdu ed. Usually the amera is
alibrated by exploitation of onstraints given by vanishing points of spa e
dire tions. To the problem of 3D re onstru tion few dierent solutions are
given whi h are des ribed below.
B.Caprile et V.Torre [CT90℄ propose a method of amera alibration
based on vanishingpoints. The proje tions of two parallel lines inthe s ene
of the amera. This fa t allows to onstru t equations useful for amera
alibration. Inthis approa hthe onstraints ome fromthe vanishingpoints
ofperpendi ulardire tions. Theprobleminsu happroa hesisthefrequently
bad numeri al onditioning inthe vanishingpoint omputation.
EdmondBoyerandRoberto Cipolla[CB98℄ propose amethodfor
re on-stru tion of urban environments based on iterative amera alibration. No
information a priori about the amera parameters is needed. Calibration is
based alsoon vanishingpoints. This approa h is basedon a linear riterion
whi h optimisesathree dimensional dire tion, the dual of a vanishingpoint
inthe3Dspa e. Todeneparalelismandperpendi ularity, uboidsareused.
Experiments show, that estimatinga spa edire tionis morerobust tonoise
perturbations than estimating an image point. This method an alibrate
severalimages. The re onstru tionisprovidedwiththe useoffew alibrated
images.
FabienVignonandEdmondBoyer[Vig99℄proposeamethodofmodelling
of urban s enes based on few images. Their method of amera alibration
is based on onstraints for amera parameters provided by the images of
uboids. The onstraints are extra ted from the proje tion matrix in ane
system oordinates atta hed to the uboid. This method of alibration
ap-plied to parallelepipeds isused in this work. The re onstru tion is based on
orrespondingpointsinfew images.
PeterSturm andStephen Maybank[SM99b℄ presentageneralalgorithm
for plane-based alibration, that an deal with arbitrary numbers of views
and alibration planes. The algorithm an simultaneously alibratedierent
views froma amerawith variableintrinsi parameters. Forsome ases they
des ribethe singularities. Experimentalresults exhibit the singularities and
reveal a good performan e in non-singular onditions. Also the proposition
of re onstru tion basedonthe knowledgeofthe metri stru ture ofplanes is
presented.
The sameauthorsin[SM99a℄propose amethodforintera tive3D
re on-stru tion of pie ewise planar obje ts from a single image. For amera
ali-brationthe methodexplainedin[CT90 ℄isused. The re onstru tionisbased
on user-provided oplanarity, perpendi ularity and parallelism onstraints.
Known intrinsi parameters of the amera allow nding proje tion rays of
points. The vanishing line of planes give information about their normal
of the sum of squared distan es between planes and points. Subsequently,
planes with unknown vanishing lines are re onstru ted in an iterative
man-ner.
A.Criminisi,I.ReidandA.Zisserman[CRZ99℄des ribeamethodof
mea-surements of distan es in the 3D s ene based on known vanishing line of a
referen e planeandthe vanishingpointofthe dire tionperpendi ulartothe
referen e plane. A new algebrai representation is developed for measuring
distan es betweenplanesparalleltothereferen eplane, omputingareaand
length ratios onplanesparallel tothe referen eplane, determiningthe
am-era position. The disadvantage of their sequential re onstru tion method is
the fa t,that errors are not spreaduniformlyin the model.
An approa h for amera alibration based on Parameterised Cuboid
Stru ture is presented in [CSC99 ℄ . This method uses a referen e
paral-lelepiped in the image to alibrate the amera and reate the oordinate
system allowing inserting virtual obje ts into the s ene. Some studies on
onstraints given by known parallelepiped parameters are provided. The
amera alibrationismadewith thequitestrongassumptionthatthe aspe t
ratio of the amerais known.
1.2 Our approa h
The aimof this work is to provide methodsfor 3D s ene re onstru tion
based on a single image. We want to extra t information of two dierent
types: geometri al information (position of points, lines, planes) and
pho-tometri (texture of surfa es). We are interested mostly in indoorand
out-doorar hite tural s enes. They usually ontain geometri alregularitiesthat
provide onstraintsuseful for amera alibrationand3Dre onstru tion.
Ex-amples are the perpendi ularity, parallelismof lines and planes and various
formsof in iden ebetweenplanes, linesand points. Ourspe ialinterestwas
to make exhaustive studies of the onstraints provided by primitives, whi h
appearveryofteninar hite turals enes: parallelepipedsandparallelograms.
The use of predened primitives whi h are ommonin s enes of interest
al-lowstominimizetheworkofauser, whointrodu esthe data,whileenabling
onsistant and a urate 3D re onstru tions.
vanishingpointsfor amera alibration, a methodbased on onstraints
pro-vided by theproje tionmatrixfromthe ane3Dspa egeneratedby uboid
stru ture tothe2D imagespa eisused. Inthebeginningwestudiedvarious
kinds of onstraintsfor amera intrinsi parametersprovided by proje tions
of orners of referen e parallelepipeds. We wanted to represent the
paral-lelepiped in a general way, allowing the optimal exploitation of ea h of its
known parameters. Some work in this area were done in [CSC99℄, but the
study was not exhaustive. Our solutionallows toderive one additional
on-straintfromeveryknownparameter. Solvingquadrati equationssystem an
do it. The amera alibration tool we developped exploits all the
informa-tion providedbyreferen eparallelepipedswhi hleadsto onstraintsoflinear
type. It gives the possibility of ombining the onstraints provided by the
proje tionsofseveralreferen e parallelepipeds. Itgivesalsothepossibilityof
using amera parameters known a priori. For example the assumption that
the prin ipalpointissituatedinthe enter oftheimageisveryreasonablein
general. Alsosometimeswe knoworwe an makeanassumption on erning
the aspe t ratio of the amera.
As se ond step weevaluated amethodof re onstru tion of the 3D
oor-dinates of points belonging to a onnex set of parallelograms. With known
ameraintrinsi parametersweare abletore onstru tthe proje tion raysof
points. Tondthe depthsofthosepointswe evaluatedamethodintrodu ed
in [CSC99 ℄ of nding the relative depths of verti es of parallelograms. We
wantedtore onstru tmanys enepointssimultaneouslytoavoidthe
in reas-ing error inthe 3D model, whi h ofteno urs in sequential approa hes.
1.3 Overview of this report
In the se ond hapter we give some denitions of proje tive geometry,
ameramodellingand alibration,and 3Dre onstru tion. Inthe third
hap-ter we introdu e the parameterisation of parallelepipeds and present
on-straints on amera intrinsi parameters provided by known parameters of
referen e parallelepipeds. We also show experimental results of tests of
ap-pli ation based on the presented ideas. In the fourth hapter we present
an algorithm for 3D re onstru tion based on sets of onnex parallelograms,
Theoreti al ba kground
Inthis hapterwepresentthetheoreti alba kgroundofthiswork. Atthe
beginningofthis hapterwegivenotations ne essarytoreadthisreport.
Af-ter this we introdu e basi denitions on erning proje tive geometry. Next
webriey des ribethe prin iples of ameramodellingand alibration. Then
we explain the on ept of 3D re onstru tion and des ribe the dieren e
be-tween the proje tive, ane and Eu lidean re onstru tion. At the end, we
briey presentthe ideaof datanormalization,introdu edin omputer vision
by Hartley, whi h an be used for numeri alstabilization of algorithms.
2.1 Notation
Bold letters represents olumn ve tors. We do not distinguish between
a point and its algebrai al representation. To represent points in 3D spa e
we use apital letters, to represent points in 2D spa e we use lower ase.
Corresponding pointsin the 3D s ene and 2D image are represented by the
same letter, e.g. Q represents the point in 3D spa e and q represents its
proje tioninthe image. Thenullve tor ofthe spa eof ndimensions willbe
represented as0
n .
Tonotematri esweuse apitalletters. Theidentitymatrixofdimension
nn is represented by I
n .
S alars, in parti ular elements of ve tors, are in lower ase, for example
we an writeQ =(q 1 ;q 2 ;q 3 ) T .
Therealspa eisrepresented byR n
. Theproje tivespa eofdimensionn
isrepresented byP
n
. Theproje tiveequalityup toas alarfa torisdenoted
as .
2.2 A few notions of proje tive geometry
2.2.1 Homogeneous oordinates
The basi mathemati al framework of our work is proje tive geometry.
The n-dimensional proje tive spa e is denoted P
n
. Points in P
n
are
rep-resented by ve tors with n+1 oordinates. They are alled homogeneous
oordinates and are dened up toa s alar fa tor:
[x 1 ;:::;x n+1 ℄ T [x 1 0 ;:::;x n+1 0 ℄ T (2.1)
if, and onlyif, 96=0 su h that
[x 1 ;:::;x n+1 ℄ T =[x 1 0 ;:::;x n+1 0 ℄ T (2.2)
Linear transformationsof a proje tive spa e are alled homographies or
ol-ineations.
2.2.2 Point at innity
The orresponden e between point oordinates in P
n and in R n is de-s ribed as 0 B B x 1 x 2 x 3 x 4 1 C C A ! 0 x 1 x 4 x 2 x4 x 3 x4 1 A (2.3)
Thistransformationisnotpossiblewhenx
4
=0. Thismightbeexplained
by the fa t that the proje tive spa e onsists of the usual ane spa e and
the plane x
4
=0, whi h is alledthe planeat innity.
To ilustrate the on ept of points at innity onsider a proje tive line.
B B x 1 x 2 x 3 1 C C A + d 0 B B x 1 x2 x 3 1 C C A + d 0 ! !inf d 0 (2.4)
Equation(??) implies,that allproje tive lines indire tiond interse t at
plane at innity in the same point [d
T ;0℄ (g. ??). 0 0 0 1 1 1 00 00 00 11 11 11 0 0 0 1 1 1 0000000000 0000000000 0000000000 0000000000 0000000000 0000000000 0000000000 0000000000 0000000000 0000000000 0000000000 0000000000 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111 1111111111 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111
P1
P2
d
d
π∼
x4=0
P=[d,0]
Figure2.1: The interse tion of parallel lines isa point atinnity.
2.2.3 The Absolute Coni
The aim of this se tion is to introdu e the on ept of absolute oni ,
whi h is useful for amera alibration. At the beginning we give brief
de-nitions of oni s and quadri s in proje tive spa e, whi hwill allowtodene
dieren es between proje tive, ane and eu lidean transformations of the
proje tive spa e. This willallow usto show the importan eof properties of
the imageof the absolute oni .
A oni in P
2
is dened as the urve on the proje tive plane onsisting
of points[x 1 ;x 2 ;x 3 ℄ T
that satisfy a quadrati equation:
3 X i;j=1 a ij x i x j =0 (2.5) A quadri of the P 3
[x 1 ;x 2 ;x 3 ;x 4
℄ that satisfyan equation of the type:
4 X i;j=1 a ij x i x j =0 (2.6)
The transformations of the proje tive spa e, whi h preserve the plane at
innity are alled the ane transformations. An important property is the
fa t thatthey preserveparallelismof linesand ratios ofdistan es onparallel
lines.
Let usdene the spe ial oni , onsisting of points[x
1 ;x 2 ;x 3 ;x 4 ℄ T with: 3 X i=1 x 2 i =x 4 =0 (2.7)
This oni is alled the absolute oni . It onsists of omplex points at
innity. The transformations of the proje tive spa e, whi h preserve the
absolute oni are alled the eu lidean transformations.
The eu lidean transformations preserve angles and ratios of lengths of
obje ts. Inotherwords,they analwaysbede omposedintorotation,
trans-lation and a global s aling. Thus the image of the absolute oni does not
depend on the position and orientation of the amera in spa e. It depends
only onthe intrinsi parameters of the amera. That is why the on ept of
the absolute oni is basi for amera alibration.
2.3 Camera modelling
Inthis se tionwedes ribesomebasi on eptsfor ameramodelling. At
the beginningwedes ribethe ameramodelusedinthiswork. Inthe se ond
part we briey present the transformations that are appliedto pointsin 3D
spa e to obtaintheir proje tions inthe image.
The ameramodelwe useisthe pinholemodel. Itisshown onthegure
??.
Thepinhole ameramodelofthe ameraisbasedontheproje tive
trans-formationofthe3Dspa etotheplane alledtheimageplane. Theproje tion
enterCisalso alledtheopti al enter. Thedistan ebetween theopti
Q
q
Optical axis
Image plane
Focal plane
Projection center
Principal
point
Figure 2.2: The pinhole amera model.
in spa e is the interse tion of the line XC and the image plane. This line
is alled the proje tion ray. The opti axis is the line passing through the
proje tion enter and that is perpendi ular to the image plane. Its
interse -tion with theimageplane is alledthe prin ipalpoint. Thefo alplaneisthe
plane ontaining the proje tion enter and parallel tothe image plane. It is
easy to observe, that proje tions of points in the fo al plane are points at
innity of the image plane.
Forthe mathemati aldes ription of thepinhole model, we onsider
nor-mally four dierent oordinates systems: a s ene system, a amera system,
a system in the image plane and a system related to the pixel geometry of
the imageplane. The relation between thosesystems isshown onthe gure
??:
The transformation between the point oordinates in the 3D s ene and
oordinatesof itsproje tioninthe imageis omposed ofthe following
trans-formations: 0 wu wv w 1 A = i r C r S s T 0 B B x y z t=1 1 C C A (2.8)
Tis thetransformationfromthe s ene oordinateframetothe amera
oor-dinate fram. It is a eu lidean transformation,so it an be de omposed into
Scene
Camera
Image
Pixels
Q
q
X
Y
Z
x
y
u
v
Figure 2.3: The pinhole ameramodel oordinatesystems.
points in the oordinates system onne ted with the amera to the retinal
plane. C is the ane transformation fromthe oordinate system onne ted
with the image plane tothe system onne ted with pixels. The omposition
of those transformations is usually represented by a proje tion matrix P
34 .
As the result wehave:
xP
34
X (2.9)
Matrix Pis dened up the s ale fa tor, so it has 34 1 = 11 degrees of
freedom.
The proje tion matrix may bede omposed asfollowsin ordertoexhibit
physi al propertiesof the amera:
P K( R j t) (2.10)
WhereR
3x3
istherotationmatrix(anorthogonalmatrix)andta"translation
ve tor" of size 3. The R and t have six degrees of freedom and they are
alled extrinsi parameters of the amera. They represent the orientation
and position of the amera with respe t to the 3D s ene. The matrix K is
the upper triangular and has thus ve degrees of freedom(it is dened up
to s ale). They orrespond tothe intrinsi parameters of the amera. On e
the intrinsi parameters of the amera are known, we say that amera is
alibrated.
K= 0 f s u 0 0 f v 0 0 0 1 1 A :
Its elements have the following meanings:
(u 0
;v
0
) - the position of the prin ipalpointsinthe image
- the aspe t ratio i.e. the ratio between horizontal and verti al dimensions of a pixel
s - the skew parameter whi h depends onthe anglebetween axes. Asnormally pixels are re tangular,in the rest of this work we will onsider s=0.
f - the fo aldistan e multiplied by height of a pixelin units of aretinal frame.
2.4 Camera alibration
As wementionedatthe endofthe previousse tion,by alibrationofthe
amera we understand the extra tionof its intrinsi parameters.
Therearetwomainapproa hestotheproblemofthe amera alibration:
Ifwe knowpositionsofpointsinspa esu h aspointsona alibration
patternwe an alibratethe amerausing oordinates oftheir
proje -tions in one or more images.Equation(??) gives ustwo onstraints on
the elements of the proje tion matrix for every known orresponden e
between a 3D point and itsimage. That implies that if we know
posi-tions inthe eu lideanspa eofatleastvepointsand onedire tion,we
an obtainthe proje tionmatrix. Then,usingthefa t,thatKisupper
triangular andRisorthogonal,we anextra t theintrinsi parameters
of the ameraby Cholesky de omposition.
Othermethodsof amera alibrationuse onstraintsprovidedbyknown
geometri al dependen ies, like perpendi ularity and parallelism of
di-re tions in the s ene. There exist several approa hes to this problem,
but theyleadusuallytosimilar onstraintsontheintrinsi parameters,
i.e. to equations based on the image of the absolute oni . Its link to
alibration and metri s ene stru ture was des ribed in se tion ??.
Todemonstrate this typeof onstraintsbriey des ribethe methodof
amera alibration based onvanishingpointsas introdu edin [CT90 ℄.
The vanishingpointof adire tionis theimageof the pointof
d
d
d
d
projection
center
vanishing
point
Figure 2.4: The vanishing point.
The main idea of the algorithm is based on the fa t, that the s alar
produ t of two perpendi ular dire tions isequal to0. Let us take two
ideal points, i.e. points at innity, [V
1 ;0℄ T and [V 2 ;0℄ T . From the
proje tions equationsv
i
KV
i
we obtain the orresponden e
V i K 1 v i (2.11)
Considering this and the fa tthat V
1 T V 2 =0 wehave: v 1 K T K 1 v 2 =0 (2.12) where the K T
is the transposition of invertion of K. The matrix
K T
K 1
isthe imageof the absolute oni .
Equation(??) implies that every pair of vanishing points of
perpen-di ular dire tions gives one onstraint on intrinsi parameters of the
amera.
This shows the utilityfor alibrationofimages of uboids. One uboid
ingeneralposition(??)bringsthree onstraintsontheintrinsi
param-eters. In many pra ti al situations some a priori information on the
intrinsi parameters is given; so the onstraints provided by a uboid
are usually su ientto alibratethe amera.
2.5 3D Re onstru tion
By the re onstru tion of the 3D model of the s ene we understand
vanishing point
vanishing point
vanishing point
cuboid
Projection center
Figure 2.5: Vanishing pointsimplied by the uboid.
obtain dierent types of re onstru tion. Usually we are interested in
re on-stru tion, whi hpreservethe anglesand atleast ratiosof lengths. This kind
of re onstru tion isnot possible if wedo not have the amera alibrated.
2.5.1 3D Proje tive re onstru tion
In the ase of the un alibrated amera we are only able to obtain a
so- alled proje tive re onstru tion. It brings some 3-dimensional information,
allowingforexampletore ognizethetypeoftheobje t(seegure??). Itwas
shown by Faugeras in[Fau92℄ and Hartleyin [HGC92℄ that itis possible to
obtain the proje tive alibration of a set of ameras without any additional
information about the s ene stru ture.
2.5.2 3D Ane and eu lidean re onstru tion
Byanere onstru tionweunderstandare onstru tion,whi hpreserves
the parallelism and the lengths ratios on parallel dire tions. To obtain this
typeofre onstru tionwehavetond thetransformationfromtheproje tive
spa e totheane spa e. Tond this transformationwehave toidentify the
plane at innity. Three points in general positions belonging to the plane
dene it. They an be obtained by use of the vanishing points (g. ??) of
three pairs of parallel lines. It was shown that using this on ept it is quite
easy to obtainthe ane re onstru tion.
model (up to the
s aleve tor).
Coplanarityandgeneralform
arepreserved.
Figure 2.6: Two threedimensionalmodels of the house.
re onstru tion, we have to extra t the intrinsi parameters of the amera.
Forthis,some additionalinformationhas tobegiven. One useful onstraint
is the onstraint of perpendi ularity ([CT90℄,[CB98 ℄). Known angles, some
knowledgeabout distan e ratios, an alsobring useful onstraints, asshown
in hapter ??
The eu lidean re onstru tion, withoutany metri informationabout the
s ene, anbeobtainedonlyuptoas alarfa tor. Thisissatisfa toryinmost
of appli ations. But it is su ient to know a single length to x the s alar
fa tor and obtainan exa t 3D re onstru tion of the s ene.
2.6 Data normalisation
One problem that we fa e almost every time when we apply ideas to
real systems is the bad onditioning of the data. Let us onsider image
points,whi harethe baseofour omputations. Theyare usuallyofthe type
x=[u;v;1℄
T
,whereu;v areofthe order10
2
;10
3
. Thetypeof onstraintsthat
of the order between 1 and 10 . The other thing is that if pixels are not
entered, their oordinates are of very dierent orders.This might auses a
bad onditioningof equation system and lead to very unstable solutions.
R.Hartleyin[Har95℄proposedanormalizationpro edureforpointssu h
as toobtain oordinates x 0 =[u 0 ;v 0 ;1℄ T ,whereu 0 ;v 0
areof theorder of1. It is
obtainedbya hangeoftheimage oordinatesysteminawaythatguarantees
that all points of interest are lying in the unit ir le. The rst step is the
translationof the enterofthe oordinatesystem tothe gravity enter ofthe
set of points. The se ond step is a s aling of oordinates to obtain lengths
of ve tors x
0
i
1.Thisidea isillustrated by g.??.
1
100
Figure 2.7: The Hartleynormalisation.
Su h transformations improve the onditionning and lead to better
Camera alibration using
parallelepipeds
In this hapterwe wouldliketointrodu e a alibration methodbasedon
parallelepipeds.
Weespe iallyare interested inurbans enes. Thebase primitiveexisting
in manmade environments isthe uboid (Fig. ??).
Figure3.1: Exampleof uboiddenition for amera alibration
The utilityof uboidstru turesindeningurbans enestru tureis
obvi-ous. Butitisalso lear,thatwe annotdenetheentires eneonlybymeans
the stru ture more interesting than uboid, but stillquiteeasyto
onstru t-the parallelepiped stru ture. We an see the examples of su h buildings on
Figures ??, ??.
(a)TorresKio,Madrid. (b) Two towers lean
towards ea h other at
anangleof15
o .
Figure 3.2: Example of parallelepiped stru ture inurbane s ene.
The se ondreason speakingfor applyingparallelepipedsisthe fa t,that
they extend the lass of obje t whi h an be used for alibration, so we an
obtain more onstraints. In the presen e of the noisy data it an
numeri- ally stabilize results. The Figure ?? shows how we an dene additional
parallelepiped stru ture, whi his not expli itlyshown onthe image.
Our methodexploitsthe onstraintsonintrinsi parametersprovided by
images of verti es of parallelepipeds in the image and known parameters of
the parallelepipeds. In order to express those onstraints, we had to hoose
a parameterization of the parallelepipeds. This is des ribed inthe rst part
of the hapter. Then we show our method for omputing an intermediate
proje tionmatrixfromtheane oordinatesystemrelatedtoparallelepiped
stru ture. Next we present onstraints on the amera intrinsi parameters
provided by olumns of this intermediate matrix. Then we des ribe our
way of using the prior knowledge onintrinsi parameters and dis uss whi h
parameters an be alibrated in degenerated ases. Finally, we des ribe our
Boston. JohnHan o ktoweris
rhomboid.
Figure 3.3: Example of parallelepiped stru ture inurbane s ene.
3.1 Parallelepiped parameterization
Our methodof amera alibration is based on mat hes between pro
je -tions oftheparallelepiped's ornersintheimageandtheir oordinatesinthe
ane spa e related tothe primitive.
Let usdene: 1 =\XY; 2 =\YZ; 3
=\ZX angles between the edges of
paral-lelepiped. 1 ; 2 ; 3 lengthsof edges
Todene onstraints onthe intrinsi parameters, we introdu ed the
param-eterization of parallelepiped shown in Figure??.
The basis hange fromane toeu lidean spa e an bedes ribed as:
E X=
4 A
bration. lelepipedstru ture.
Figure3.4: The use of parallelepipeds for additional onstraints.
where
4
is the matrix of hangeof the base and an be expressed as
4 = 0 B B x 1 x 2 x 3 0 0 y 2 y 3 0 0 0 z 3 0 0 0 0 1 1 C C A (3.2) where x 1 = 1 (3.3) x 2 = 2 os 1 (3.4) y 2 = 2 sin 1 (3.5) x 3 = 3 os 3 (3.6) y 3 = 3 os 2 os 3 os 1 sin 1 (3.7) z 3 = 3 s 1 ( os 3 ) 2 os 2 os 3 os 1 sin 1 2 (3.8)
3.2 Computation of the intermediate
proje -tion matrix
Let us x now the oordinates of a parallelepiped's verti es inan ane
y
x
z
x’
y’
z’
(0,0,0)
(x1,0,0)
(x2,y2,0)
(x3,y3,z3)
λ3
λ2
λ1
θ3
θ2
θ1
Figure3.5: The ane oordinatesystem relatedto the parallelepiped.
x’
y’
z’
(1,1,0)
(1,0,0)
(0,0,0)
(0,1,0)
(0,0,1)
(0,1,1)
(1,1,1)
(1,0,1)
Figure 3.6: The xed oordinates of verti es inthe ane oordinatesystem
i
parallelepiped and q
i
itsproje tion in the image. From the previous se tion
we know that the oordinates of the point Q
i
in the eu lidean oordinate
systemareexpressed by
4 Q
i
. LetPbetheproje tionmatrixfromeu lidean
spa e to the image. We havethe following dependen e:
P 4 Q i =q i (3.9)
Every dependen e of the type ??(i.e. every known proje tion of one
of the parallelepiped's verti es ) gives us the following onstraints on the
matrix M=P 4 : 0 wu wv w 1 A = 0 m 11 m 12 m 13 m 14 m 21 m 22 m 23 m 24 m 31 m 32 m 33 m 34 =1 1 A 0 B B x y z t=1 1 C C A (3.10)
This allows the onstru tion of two linear equations from every
depen-den e: m 31 xu+m 32 yu+m 33 zu+u=m 11 x+m 12 y+m 13 z+m 14 m 31 xv +m 32 yv+m 33 zv+v =m 21 x+m 22 y+m 23 z+m 24 (3.11)
whi h leads tothe following equationsystem of the type: AX=B
0 B x 1 y 1 z 1 1 0 0 0 0 x 1 u 1 y 1 u 1 z 1 u 1 0 0 0 0 0 x 1 y 1 z 1 1 x 1 v 1 y 1 v 1 z 1 v 1 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 C A 0 B B B B B B B B B B B B B B B B B B m 11 m 12 m 13 m 14 m 21 m 22 m 23 m 24 m 31 m 32 m 33 m 34 =1 1 C C C C C C C C C C C C C C C C C C A = 0 B u 1 v 1 . . . 1 C A (3.12)
depends on the rank of A.
if rk A =11 one unique solution,up toa s alarfa tor, isobtained
si rk A<11an innitefamily of solutionsis obtained.
This impliesthatweneedatleastsixpointsingeneralpositiontoobtain
thesolution. Intheoryweneedonlyvepointsandonedire tion(adire tion
provides one onstraint) to obtain the matrix. In pra ti e we an usually
identifyatleastsixproje tionsoftheprimitive'sverti es. Ontheotherhand,
to improve the results, we usually use over onstrained systems. We nd the
solutionof su ha systemby minimizationof theexpression kAX Bkinthe
sense of the least squares using the SVD method (des ribed for example in
[PTVn92 ℄.
3.3 Constraints on the intrinsi parameters
In this se tion we will des ribe the onstraints for the amera intrinsi
parameters provided by matrix P
4
introdu edinthe previous se tion.
Using the equation (??) we an write
P 4
K( R j t)
4
(3.13)
Let us onsider the matrix X
33 =KR,where = 0 x 1 x 2 x 3 0 y 2 y 3 0 0 z 3 1 A (3.14)
Let usevaluate the expression X
T K
T K
1
X using the fa tthat R is the
orthogonal matrix. Weobtain:
X T K T K 1 X =(KR) T K T K 1 KR= T R T K T K T K 1 KR= T (3.15) The matri es T and !K T K 1
T 2 1 1 2 os 1 1 3 os 3 1 2 os 1 2 2 2 3 os 2 1 3 os 3 2 3 os 2 2 3 A (3.16) ! 0 1 0 u 0 0 2 2 v 0 u 0 2 v 0 2 f 2 +u 2 0 + 2 v 2 0 1 A : (3.17)
Weare lookingforelements of! onlyup toas alarfa tor, sowe anx
one of its elements, for example !
11
. Considering equations (??), (??) and
(??) we obtain: X T 0 ! 11 =1 0 ! 13 0 ! 22 ! 23 ! 13 ! 23 ! 33 1 A X 0 2 1 1 2 os 1 1 3 os 3 1 2 os 1 2 2 2 3 os 2 1 3 os 3 2 3 os 2 2 3 1 A (3.18)
Equation (??) allows to make some observations about what kind of
onstraints on intrinsi parameters we an obtain depending on the kind
of information about parallelepiped parameters. Considering that
T
is
known up to as alar fa tor, we have:
One known angle
ij
= 90
o
gives one linear onstraint on elements of
!: X i T !X j =0 (3.19)
One known angle
ij
givesone nonlinear onstraint onelementsof !:
os 2 ij = (X i T !X j ) 2 (X i T !X i )(X j T !X j ) (3.20)
One known length ratio r
ij =
i
j
gives one linear onstraint on !
ele-ments: X i T !X i (r ij ) 2 X j T !X j =0 (3.21)
then up tove onstraints onthe elementsof !: X i T !X j ( T ) ij ( T ) kl X k T !X l =0 (3.22)
If more parallelepipedsare available, we justin lude new equations into
an equation system. At the beginningwe add only onstraints of the linear
type whi h leads to an equation system of type AX = 0. Having at least
four equations we an obtain a unique solution for the ve tor of unknowns
[! 13 ;! 22 ;! 23 ;! 33 ℄ T
using the SVD method. If the linear system is
under- onstrained we an use non-linear onstraints to limitthe spa e of solutions
obtained by applying the SVD method for under onstrained equation
sys-tem.
On e the elements of the matrix ! are known the intrinsi amera
pa-rameters an be extra ted via [SM99b℄:
2 = ! 22 ! 11 (3.23) u 0 = ! 13 ! 11 (3.24) v 0 = ! 23 ! 22 (3.25) f 2 = ! 11 ! 22 ! 33 ! 22 ! 2 13 ! 11 ! 2 23 ! 11 ! 2 22 (3.26)
3.4 Prior knowledge about amera parameters,
degenerated ases
Veryoftenweknoworwe anmakeassumptions on erningsome
intrin-si parameters of the amera. Formost ameras, for example, the prin ipal
pointliesinthe enter of the image. Information aboutthe aspe t ratio an
usually be found in a amera's te hni al do umentation, or an sometimes
be assumed to be unit. The fo al distan e is the only parameter that we
always allow toextra t all the intrinsi parameters. It is easy to show both
algebrai allyand geometri allythat the following degenerated ases exist:
Ifonefa eoftheprimitiveisparalleltotheimageplaneanditslengths
areunknownthenedgesofthisfa egiveno onstraintsontheprin ipal
point,fo aldistan e and skew fa tor( Eq.?? isalways true).
The dire tion perpendi ular to the image plane allows to lo alize the
prin ipalpoint,but doesnot onstrain otherintrinsi parameters.
Those degenerated ases are illustrated by Figure ??
C
Figure3.7: Adegenerated ase. Proje tionsoflineslyingintheplaneparallel
to the image plane do not onstrain the intrinsi parameters if lengths are
unknown. The proje tion of the perpendi ular tothe image plane dire tion
determines the prin ipalpoint.
There are two ways to use the prior knowledge about the intrinsi
pa-rameters.
0nemethodistotranslatetheimage oordinatesystemtotheprin ipal
point and to s ale its axes su h that in further al ulus u
0
= 0;v
0 =
0; =1.
The se ond solutionisto add tothe mainequation system, depending
13 0 ! 22 = 2 (3.28) v 0 ! 22 +! 23 = 0 (3.29) 3.5 Experimental results
Toverifyrobustnessofourmethodswemadefewtests. Atthebeginning
wetestedour alibrationmethodonaseriesof syntheti data. Inthe se ond
stage wemadetest basedonphotoofthe alibrationpattern. Atthe endwe
tested our appli ation on real data. In this se tionwe des ribe our results.
3.5.1 Syntheti data
Toobtaingeneraldedu tions on erningour methodweappliedtests on
syntheti data. We appliedthe same test totwo dierent primitives:
First test was based on parallelepiped with angles (30
0 ;45 0 ;60 0 ) and lengths given.
Se ond test wasbased onparallelepiped withangles(60
0 ;90 0 ;90 0 )and unknown lengths.
Our appli ation have few types of parameters. First type are image
oordinates of primitives proje tions. The se ond type are known a priori
intrinsi parameters like position of a prin ipal point and a skew ratio. In
order toverify the sensitivity ofour methodto thoseparameters we applied
tests of the method with noisydata.
On the other side, we wanted to verify the behavior of the method in
degenerated ases. Inordertodoit,wetestedresultsof alibrationbasedon
parallelepiped rotated aroundthe verti al axisparallel tothe imageplane.
The alibration results given above were obtained as a median from a
seriesofrandomtest. Ea htestwasprovidedwithandwithoutdata
normal-izationinthepre omputationstage. Weassumedthatimagesizeis512512
and theprin ipalpointliesinthe enteroftheimage. Weformulatedseveral
but in the ase of our method the improvement of results is not very
signi ant. This an be seen on Figures ?? and ??. They present
results of alibration with and without normalization for both tested
primitives rotated aroundthe verti al axis. We an see that positions
lose todegenerated asesdoesnot allowto alibratea amerawith or
withoutnormalization. Alsoresultsof alibrationarenotverydierent.
Similardedu tionsweremadeintestsofsensitivitytonoiseonprogram
parameters.
Knowledge about lengthsof parallelepiped, thanks toadditional
equa-tions to the alibration system, improves obtained results in tests on
sensitivity to noiseon known apriori intrinsi ameraparameters and
test onrotatedprimitives.
Known lengths of edges allows to alibrate also in otherwise
degener-ated ases (des ribed in se tion ??). Figures ?? and ?? show results
of alibration of a parallelepiped, whi h is rotated around the
verti- al axis parallel to the image plane. The random noise of two pixels
was applied to primitivesverti es image oordinates. In ase of
paral-lelepiped, whi h lengths are not known, every time its fa e is parallel
to the image plane, alibration is not possible. In ase of known edges
in su h ases wesee onlyin reasing error.
In the se ond test we applied random noise onimage oordinations of
primitivesverti es proje tions.
Calibration in ase of primitive whose edges lengths are given is
pos-sible up to 3 pixels noise (Figure ??). In the se ond test results are
reasonable up to six pixels noise (Figure ??). It an be explained by
the fa t, that in ase of the rst primitive (with allparameters given)
weobtainmoreequationsbasedonnoiseddatathaninthese ondtest.
It may ause worst results.
The third test (Figures ?? and ??) wasapplied to verify the inuen e
of the error on position of a prin ipal point in the image. We tested
behavior of the method with distan e between (u
0
;v
0
) real and given
as a program parameter form 0 to 200 pixels. Experiments on both
primitives show, that the method is not very sensitive for not exa t
0
2
4
6
8
10
12
14
16
18
650
700
750
800
850
900
950
Rotation angle
f
simple
normalized
f theor.
(a)Test1,rotation
0
2
4
6
8
10
12
14
16
18
650
700
750
800
850
900
950
Rotation angle
f
simple
normalized
f theor.
(b)Test 2,rotationFigure3.8: Test of alibrationbasedonparallelepipedrotedaroundaxis
par-allel to the image plane for primitive with known lengths(a) and unknown
lengths (b). In both ases degenerated ases are seen, but in ase (a)
0
2
4
6
8
10
12
14
16
18
400
500
600
700
800
900
1000
1100
1200
Rotation angle
f
(a)Test1,rotation
0
2
4
6
8
10
12
14
16
18
400
500
600
700
800
900
1000
1100
1200
Rotation angle
f
(b)Test 2,rotation0
1
2
3
4
5
6
7
8
9
10
11
400
500
600
700
800
900
1000
1100
1200
Image coordinates
f
( )Test1,noisedimage oordinates
0
1
2
3
4
5
6
7
8
9
10
11
400
500
600
700
800
900
1000
1100
1200
Image coordinates
f
(d)Test2,noisedimage oordinates
Figure 3.9: Graphs of averages and standard deviations from tests on
syn-theti data.(a)- test of alibration based on primitive with angle between
fa e and image plane between 0 180, edges lengths are known; (b)- test
of alibration based on primitive with angle between fa e and image plane
between 0 180, edges lengths are unknown;( )- test with error on image
oordinates from0 10 pixels, edges lengthsare known;(d)- test with error
0
20
40
60
80
100
120
140
160
180
200
220
400
500
600
700
800
900
1000
1100
1200
Position of a principal point
f
(a)Test1,noisedposition (u0,v0)
0
20
40
60
80
100
120
140
160
180
200
220
400
500
600
700
800
900
1000
1100
1200
Position of a principal point
f
(b)Test2,noisedposition(u0,v0)
0
0.2
0.4
0.6
0.8
1
400
500
600
700
800
900
1000
1100
1200
Skew ratio
f
( ) Test 1,noisedskew ratio
0
0.2
0.4
0.6
0.8
1
400
500
600
700
800
900
1000
1100
1200
Skew ratio
f
(d)Test 1,noisedskew ratio
Figure 3.10: Graphs of averages and standard deviations from tests on
syn-theti data.(a)- test with error on position of prin ipal point, edges lengths
are known; (b)- test with error onposition of prin ipalpoint,edges lengths
are unknown;( )-test witherroronskewratio,edgeslengthsare known
then 20%. This resultshows, that assumption that theprin ipalpoint
lies inthe enter of the image is usually very reasonable and even if it
is false, alibrationshould give quitegoodresults.
Prior knowledge about the aspe t ratio improves the alibration. On
the other hand if the value given is not exa t, it spoils results. It is
shown onFigures?? and ??.
3.5.2 Calibration pattern image
(a)Familyof uboids. (b)Familyofparallelepipeds.
Figure3.11: Experiment with alibrationpattern.
For the se ond test we used the photo of a alibration pattern. We
examined two families of primitives: uboids(Fig. ??) and parallelepipeds
(Fig. ??)withdierentsizeofedges. Weappliedalsoteststo ompareresults
for every family with known and unknown lengths of edges .In all ases we
assumed that the prin ipal point lies in the enter of the image. Thanks to
this test we made more dedu tions about dependen e of alibration results
on orientation and size of the primitive. Realvalues of intrinsi parameters
where omputedbystandardmethodusingknownpositionsof ontrolpoints
of the alibration patterninthe spa e.
Intables??and??wegiveexemplary alibrationresultsforprimitives
other hand we observe, that primitivestoo big, so with proje tions of
verti es losetobordersoftheimage,giveworstresultsthanprimitives
well entered. Wegiveexemplaryresultsforprimitiveswithknownand
unknown lengths of edges.
lenght 1 4 5 6
uboids 49.5 5.6 6.2 3.7
uboids+lengths 47.4 3.2 3.6 1.6
parallelepipedes - 3.0 2.5 0.1
parallelepipedes +lengths - 2.7 2.4 0.4
Table 3.1: Relativeerrors for f in %
lenght 1 4 5 6
uboids 14.6 3.7 2.9 2.8
uboids+lengths 9.7 2.2 2.8 2.0
parallelepipedes - 0.0 2.4 2.2
parallelepipedes+lengths - 1.1 1.1 0.0
Table 3.2: Relativeerrors for in %
Calibrationisinuen edbyexisten eofsquaredfa es. Figures??show
results. Three harts are made for three dierent values of the length
of the edge (l
3
) of the uboid. On the x axis are values of length l
2 .
The y axis presents results of fo al omputationfor dierent values of
l 2
. We an observe, that the best results are not for the biggest value
of l
2
,but forthe value losest tothe lengthl
3 .
3.5.3 Real s enes
The most interesting inour work were tests onimages of real s enes. In
this ase alibration results are not really the most signi ant. In pra ti e
wedonot knowintrinsi parametersof a amera,with whi h the imagewas
1
2
3
4
5
6
7
1250
1300
1350
1400
1450
l1
f
l3=4
f theor.
l2=3
l2=4
l2=5
l2=6
l2=7
(a)l 3 =31
2
3
4
5
6
7
1250
1300
1350
1400
1450
l1
f
l3=5
f theor.
l2=3
l2=4
l2=5
l2=6
l2=7
(b)l 3 =41
2
3
4
5
6
7
1250
1300
1350
1400
1450
l1
f
l3=6
f theor.
l2=3
l2=4
l2=5
l2=6
l2=7
( )l 3 =5We will des ribe then s enes, whi h had been re onstru ted and problems
Constru tion of textured 3D
model
The aimof this proje t was to onstru t 3D models of s enes from one
image. The onstru tion of models ontains two problems: lo alization of
model points in the 3D spa e and photo realisti rendering of surfa es. In
this hapter wedes ribeour approa hto these problems.
There are many approa hes to the onstru tion of 3D models. Having
alibratedthe amera,we aneasilyobtainproje tionraysalongwhi hmodel
points are situated. The problem of 3D re onstru tion lies in nding exa t
positions of points on proje tion rays. Most of the resear h that has been
done in this domain uses two or more images to obtain positions of model
points in spa e. For single images Peter Sturm and Stephen Maybank in
[SM99a℄ propose a method based on sets of planes and sets of pointslinked
together. Antonio Criminisi and Andrew Zisserman proposed another
ap-proa h in [CRZ99℄. It uses measurements onplanes and lines for sequential
re onstru tion of 3D oordinates of points.
Inourworkweusedtheideaintrodu edin[CSC99 ℄for omputingrelative
depths of parallelepiped's orners. Weevaluateda methodof re onstru tion
of s enes onsisting of a set of onne ted parallelograms,whi h is des ribed
in the rst se tion.
The laststage of the proje t wasthe extra tionof photometri
informa-tion. Textures obtained by simple interpolation of olors on surfa es have
of models.
Finally,wepresentresultsofsomere onstru tionsfromimagesofoutdoor
and indoors enes.
4.1 3D points re onstru tion
We assume that the amera is alibrated and so we know the matrix K
of intrinsi ameraparameters. This enablesus toba kproje t imagepoints
to 3D along their proje tion rays. The problem of re onstru tion is now to
nd their exa t positions onproje tion rays.
A 3D pointwhose image isgiven by qhas oordinates
Q= q 0 1 (4.1) where q 0 K 1 q and jjq 0 jj = 1. q 0
is alled the viewing dire tion. The
unknown expresses the distan e of Q form the opti al enter and hen e
denes its position onthe proje tion ray (Fig. ??).
000
111
000
111
000
000
111
111
000
111
000
000
111
111
000
000
111
111
00
11
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
00000000000000000000000000000000000000000000000
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
11111111111111111111111111111111111111111111111
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
000000000000000000000000
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
111111111111111111111111
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
000000000000000000000000000000000
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
111111111111111111111111111111111
000000000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000000000
111111111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111111111
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
0000000000000000000
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
1111111111111111111
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
000000000000000000000000000000000000000000000000000000
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
111111111111111111111111111111111111111111111111111111
q’
q
Q
λ
projection center
image plane
Figure4.1: Viewing dire tions of image points.
For every parallelogram inspa e dened asshown inFigure ?? we have
Q4
Q3
Q2
Q1
Figure4.2: Constrainton 3D positions of parallelogramverti es.
Q 1 Q 2 +Q 3 Q 4 = 0 (4.2)
This impliesthat dire tions q
0
i
the parallelogram'sverti esprovidethree
equations for points depths
1 ; 2 ; 3 ; 4 : 1 0 q 0 11 q 0 12 q 0 13 1 A 2 0 q 0 21 q 0 22 q 0 23 1 A + 3 0 q 0 31 q 0 32 q 0 33 1 A 4 0 q 0 41 q 0 42 q 0 43 1 A =0 (4.3)
If we x one depth,we obtain remainingthree depths. Also,for a set of
parallelograms onne tedbyatleastonepoint,itissu ienttosetarbitrarily
one depth to ompute the positions of all verti es of parallelogram. If we
know at least one length in the s ene we an x the depths to obtain the
eu lidean modelpreserving not onlyratiosof lengths,butalsotheirabsolute
values.
In ourmethodwe reateatwodire tionalgraph,whosenodes arepoints
of parallelograms and onne tions are edges of parallelograms. Next we
hoose thelargest onnexsubsetof thegraphand weenumeratenodes. Now
equa-0 B B B B B q 0 11 q 0 21 q 0 31 q 0 41 0 0 0 0 ::: q 0 12 q 0 22 q 0 32 q 0 42 0 0 0 0 ::: q 0 13 q 0 23 q 0 33 q 0 43 0 0 0 0 ::: 0 0 0 q 0 41 q 0 51 q 0 61 q 0 71 0 ::: . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 C C C C C A 0 B B B B B B B B B B B B B B B 1 2 3 4 5 . . . . . . . . . . . . 1 C C C C C C C C C C C C C C C A = 0 B B B 0 0 . . . . . . 1 C C C A (4.4)
Afterxing onedepththe resolutionoftheequationsystemabovebrings
us informationabout depths of allpointsbelonging tothe set.
4.2 How to use photometri data for photo
re-alisti rendering
The last stage of 3D modela quisition is the extra tion of photometri
information. For this purpose we divided re onstru ted parallelograms into
triangles, adding triangles not belonging toany parallelogram, but with
re- onstru ted verti es. Textures obtained by simple interpolation of olors in
triangles in the image and trianglesin the modelare deformed. This is due
to the fa t, that the transformation from the s ene obje t to the image is
perspe tive, and the transformation from the triangle in the image to the
triangle in the 3D model is linear. Thus without any orre tion we an not
re onstru t the real pattern.
There are few approa hes tothis problem. We an for example
pre om-pute textures for ea h fa e to obtain the view as taken by a amera whose
opti al axis is parallel to the fa e's normal. In order to do this we have to
ompute the rotationbetween the opti al axis(in this work weassume that
it is the z-axis) and the fa e's normal. This allows to nd the homography
between the polygon onthe imageand polygon onnew texture.
Toobtain photorealisti renderings we applied amethodof texture
proje -The dieren ebetween rendering with and without the perspe tive
or-re tion is shown onthe Figure??.
(a) Model withtexture
or-re tion.
(b) Model without texture
orre tion.
Figure 4.3: Example of texture orre tion.
4.3 Experimental results
Totest our methods, we onstru ted several models from photos of real
s enes. We used both outdoor and indoor s enes. In this se tion we shortly
des ribe hara teristi sof every s ene, give examples of problems whi h an
o urduringre onstru tionandpresentourresults. Ea hre onstru tionwas
made with the assumption, that the prin ipal point liesin the enter of the
image.
Our rst re onstru tion was the Sioux building. It is an example of a
simple ountry house. If we do not need very detailed re onstru tion
we an dene main obje t by only two primitives. It means that
re- onstru tion is very qui k. As is shown on the Figure ?? angles are
on the modelredu e this ee t.
The Notre-Dame pla e in Grenoble. It was dened as uboid (the
tower) and set of parallelograms. During denition of the model we
had todenevery arefulythefa ebetween thetowerandhouses. It is
very narrow, soquitesensitive toin orre t data. These ond di ulty
was the fa tthat anglesare not really straight inthe real s ene- walls
of houses are not really at, and also fronts of houses are not really
paralleltothefrontofthe tower. Butasinthepreviousre onstru tion
wesee(Fig. ??),thatre onstru tedanglesarequite losetothereality.
The photo was taken by a amera that is known to have unit aspe t
ratio. The value ofthe aspe t ratio obtainedbyour methodis0:97, so
very lose tothe real one.
Torres Kio from Puerta Europa in Madrid. This is an example of a
modernstru ture whi h annotbedenedby a uboid. For alibration
we used only two onstraints provided by two pairs of perpendi ular
dire tions. Therelativeerrorobtainedforre onstru tedangleswasless
than 3%, and less then2% for lengths'ratios. Some views ofthe s ene
are shown onFigure ??.
The indoor s ene of the room under the roof (Fig. ??). To dene
ontrolpointsof uboids weused ontrolmarks. Thisphotowastaken
with thesmallestfo alofthe amera. Inresult,we an seethe ee tof
distortiononthe image. It ausedseveral problems with olinearityof
pointsand oplanarityoffa esinre onstru tedmodel(Figure??). But
general geometry of the s ene is ni ely re onstru ted. Also textures,
even onfa esnot verywellseen by the amera,are wellre onstru ted.
Generally we an see on gures above, that angles are usually well
re on-stru ted. What is also important, that we an obtain good slide of the
( ) (d)
Figure 4.4: (a)-the original image of the Sioux building; (b)-( )-(d)-photos
( ) (d)
Figure 4.5: (a)-the original image of Notre Dame pla e; (b)-( )-(d)- photos
( ) (d)
( ) (d)
Figure 4.7: (a)-the original image of indoor s ene;(b)-( )-(d)-photos from
Con lusion
Inthisworkwehaveapplied omputervisiontheorytodevelopamethod
of 3D model a quisition from single images. We were interested mostly in
outdoorand indoor urbans enes.
In aseofasingleimage,weneedadditionalinformationaboutthe
topol-ogy of the s ene to be able to alibrate the amera and then onstru t the
eu lidean 3D model. In our approa h, to obtain the ne essary data, we
employed primitivesappearinginar hite tural environments. The most
fre-quent and very useful primitivesare parallelepipeds and parallelograms.
At the rst stage we studied onstraints on ameraintrinsi parameters
provided by prior knowledge about parameters of parallelepipeds appearing
in the s ene. The method of amera alibration was based on the method
of alibration based on uboids proposed by Fabien Vignon and Edmond
Boyerin[Vig99℄. Our methodallows usingnot only onstraintsprovided by
perpendi ular dire tions, but all the information about lengths and angles
of parallelepipeds. One ompletely dened primitive allows for omplete
alibrationofthe amera. Itpermitsalsoto alibratethe amerainotherwise
degenerated ases. Experimental tests of alibration brought good results
even in the presen e of noisydata.
In the se ond stage we developed a method for re onstru ting
three-dimensional models based on sets of onne ted parallelograms. It o urs
that su h stru tures are su ient to dene many urban s enes. Results of
re onstru tions of models from photos of real s enes demonstrate good
pendent modules: alibration, re onstru tion and modelvisualization. The
module for photorealisti rendering of the three-dimensional model was
im-plemented using OpenGL.
There are several extensions that may be added to our method. For
ex-ample other primitives an be added, like uboids with triangle prisms on
the top, spheres or ylinders. On the other hand an intera tive system
as-sisting indenition of additionalprimitivesinpartiallyre onstru ted s enes
ould be developed. Also, intera tive methods for ombined re onstru tion
of twonon interse ting sets of onne ted parallelograms ouldbedeveloped.
Interesting results ould be obtained by joining models re onstru ted from
photos of dierent parts of a s ene. Dependen ies provided by known
rela-tive positions between primitives and proje tions of one primitiveinseveral