This action might not be possible to undo. Are you sure you want to continue?

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

**Andrei Zaharescu, Radu Horaud, R´ emi Ronfard and Loic Lefort
**

INRIA Rhone-Alpes, 655, Avenue de l’Europe, 38330 Montbonnot, France

{andrei.zaharescu, radu.horaud, remi.ronfard, loic.lefort}@inrialpes.fr

Abstract

In this paper we address the problemof recovering structure

and motion from a large number of intrinsically calibrated

perspective cameras. We describe a method that combines

(1) weak-perspective reconstruction in the presence of noisy

and missing data and (2) an algorithm that updates weak-

perspective reconstruction to perspective reconstruction by

incrementally estimating the projective depths. The method

also solves for the reversal ambiguity associated with afﬁne

factorization techniques. The method has been successfully

applied to the problem of calibrating the external param-

eters (position and orientation) of several multiple-camera

setups. Results obtained with synthetic and experimental

data compare favourably with results obtained with non-

linear minimization such as bundle adjustment.

1. Introduction

The problem of 3-D reconstruction using afﬁne factoriza-

tion has received a lot of attention. The method initially

introduced by Tomasi & Kanade [1] and subsequently elab-

orated by many others – see section 1.2 – is attractive for

a number of reasons: (i) it provides a closed-form solution,

(ii) it can deal with noisy, uncertain, erroneous, and missing

data, (iii) and it is upgradable to Euclidean reconstruction.

Although initially designed to solve for rigid shape, it has

been succesfully extended to deformable shape as well as

to multiple rigid shapes (motion segmentation). Neverthe-

less, afﬁne factorization has a number of limitations as well.

First, it uses weak-perspective or paraperspective camera

models which are approximations of the perspective model.

Second, it recovers orientation and shape up to a reversal

ambiguity. Third, it recovers translation and shape up to a

scale ambiguity.

The problem of multiple-camera calibration has received

a lot of attention in the past. This problem is two-fold: esti-

mate both the intrinsic (internal camera parameters) and the

external parameters (rotation and translation) with respect

to a global reference frame. Intrinsic parameters can be es-

timated independently for each camera and several software

packages are available today. However, it is more com-

plex and more tedious to estimate the extrinsic parameters,

especially when a very large number of cameras must be

calibrated – thirty or more cameras. In practice, the cam-

eras must simulteneously observe a large number of points.

Point correspondences must be established accross the im-

ages under a multiple wide-baseline setup. Based on these

point matches, both the rotation/translation parameters and

the 3-D point coordinates must be recovered. The method

of choice for multiple-camera calibration is bundle adjust-

ment [2], [3]. It is a non linear minimization technique that

requires proper initialization.

In this paper we propose a robust incremental perspec-

tive factorization method that features the following prop-

erties:

• It can deal with noisy and missing 2-D data;

• It provides accurate Euclidean reconstruction by using

a perspective camera model;

• It solves for the orientation/shape reversal ambiguity;

• It exploits the motion parameters – recovered in

closed form through the factorization process – to pro-

vide a solution for the multiple-camera external cali-

bration problem, and

• It compares favourably with the results obtained with

ground-truth methods such as bundle adjustment.

Paper organization. The remaining of this paper is orga-

nized as follows. In section 1.1 we recall the shape-and-

motion recovery problem using a factorization formulation

and, in particular, we establish the link between full per-

spective and weak-perspective. Section 1.2 brieﬂy describes

previous work with emphasis on robust afﬁne factorization

and on several attempts to deal with the projective case.

Section 2 describes the robust method that we propose for

incrementally solving for perspective factorization. Section

3 describes a practical method for estimating the external

parameters of a multiple-camera setup. We show results

1

(a) Partial view of a 30-camera setup (b) Calibration data gathered by moving a

stick

(c) External multi-camera calibration

Figure 1: Outline of the multiple-camera calibration method described in this paper.

obtained with 30 cameras in two different layouts. We anal-

yse the performance of our method in the presence of noise

and of missing data.

1.1. Problem formulation

We consider a calibrated perspective camera model in

which case the relationship between 3-D points and image

points expressed in camera coordinates can be written as:

x

ij

=

r

x

i

· P

j

+ t

x

i

r

z

i

· P

j

+ t

z

i

=

R

x

i

· P

j

+ x

0

i

1 + ε

ij

(1)

y

ij

=

r

y

i

· P

j

+ t

y

i

r

z

i

· P

j

+ t

z

i

=

R

y

i

· P

j

+ y

0

i

1 + ε

ij

(2)

where a 3-D point P

j

is projected onto the i-th camera at

location s

ij

. The rotation matrix r

i

= [r

x

i

r

y

i

r

z

i

] and the

translation vector t

i

= (t

x

i

t

y

i

t

z

i

) correspond to the external

camera parameters. Dividing the above equations with the

depth t

z

i

we obtain a similar set of scaled equations. The

following notations have been introduced: R

x

i

= r

x

i

/t

z

i

,

R

y

i

= r

y

i

/t

z

i

. Moreover, x

0

i

= t

x

i

/t

z

i

and y

0

i

= t

y

i

/t

z

i

are

the camera coordinates of a point s

0

i

which is the projection

of P

0

(the origin of the 3-D frame). Also, ε

ij

denotes the

following ratio:

ε

ij

=

r

z

i

· P

j

t

z

i

(3)

We also introduce the perspective depth λ:

λ

ij

= 1 + ε

ij

(4)

The perspective projection equations (1) and (2) can now be

rewritten as:

λ

ij

s

ij

−s

0

i

=

¸

R

x

i

R

x

i

P

j

(5)

For k cameras and n points, i ∈ {1 . . . k}, j ∈ {1 . . . n} one

can write:

S

λ

= MP (6)

So far we considered a perspective camera model. If an

afﬁne camera model is used instead, the model above be-

comes simpler. In particular, under weak perspective λ

ij

=

1, ∀(i, j) and we have:

S = MP (7)

1.2. Previous work

There are several ways for solving the afﬁne factorization

problem, i.e., eq. (7). Initially it was proposed to minimize

the Frobenius norm by Tomasi & Kanade [1]:

min

M,P

S −MP

2

F

(8)

In the absence of noise, the singular value decomposi-

tion of the measurement matrix provides an optimal solu-

tion. The latter is deﬁned up to an afﬁne transformation T,

MP = MTT

−1

P, that can be determined using Euclidean

constraints. It is worthwhile to notice that both T and −T

can be used. This sign ambiguity corresponds to the well

known reversal ambiguity when the 3-D world is observed

with afﬁne cameras.

Anandan & Irani [4] extended the classical SVD ap-

proach to deal with the case of directional uncertainty. They

use the Mahalanobis norm instead of the Frobenius norm.

They reformulate the factorization problem such that the

Mahalanobis norm can be transformed into a Frobenius

norm. Finally they develop a modiﬁed factorization tech-

nique. More generally, a central idea is to introduce a

weight matrix W of the same size as the measurement ma-

trix S. The Frobenius norm becomes in this case:

min

M,P

W⊗(S −MP)

2

F

(9)

where ⊗ is the Hadamard product: A = B ⊗ C ⇐⇒

a

ij

= b

ij

c

ij

. The weights w

ij

describe the conﬁdence that

2

we have in an image datum and the values of these weights

vary from 1 (highly conﬁdent data) to 0 (missing data). The

most common way of minimizing eq. (9) is to use alterna-

tion methods: these methods are based on the observation

that if either M or P is known, there is a closed-form solu-

tion for the other that minimizes eq. (9). The PowerFactor-

ization method [5], [6], [7] as well as many other methods

[8], [9], [10] fall into this category. PowerFactorization [5]

presents a number of attractive features: It converges by

starting with any random input, it requires a lot less com-

putation than SVD (5% of the time for a 500 ×500 matrix)

and it deals well with missing data (60-90%). Another way

to alternate between motion and shape estimation is to use

the EM algorithm and a nice formulation that uses factor

analysis is provided by Gruber and Weiss et al. [11], [12].

Alternatively, robustness both to noisy data and to out-

liers can be achieved through adaptive weighting by itera-

tively re-estimating the weight matrix W which amounts to

iteratively modify the data S, [13]. The method proposed

in [13] uses eq. (8) in conjunction with a robust weighting

matrix (truncated quadratic or Huber’s M-estimator) to iter-

atively approximate eq. (9), getting a temporary optimum.

The approximation is performed by modifying the original

data S such that the solution to eq. (8) with modiﬁed data

˜

S

is the same as eq. (9) with the original data:

min

M,P

W⊗(S −MP)

2

F

= min

M,P

˜

S −MP

2

F

(10)

The weights are updated at each iteration, using an M-

estimator. This method may well be viewed as both an

iterative and an alternation method because the matrix M

is estimated using SVD, than the P is estimated knowing

M, and ﬁnally the image residuals are calculated and the

data (the weights) are modiﬁed. Buchanan & Fitzgibbon

[14] propose to solve eq. (9) through non-linear minimiza-

tion over the motion and shape parameters. They review

the alternation methods and they compare their speed of

convergence with a damped Newton method (a variation of

the Levenberg-Marquardt method). They propose a hybrid

method that switches from alternation to damped Newton.

The factorization methods mentioned so far can only

deal with afﬁne camera models. The latters are not well

suited for accurate estimation of motion and shape parame-

ters. It is more difﬁcult to solve for the projective multiple-

camera calibration and/or reconstruction problem. The

most accurate results are obtained through bundle adjust-

ment methods [2], [3] which require proper initialization.

Sturm and Triggs [15] introduced the use of factorization in

multiple image projective structure from motion. Their for-

mulation is similar in spirit to eq. (6) and their method relies

on fundamental matrices to solve for the projective depths.

Subsequently, Heyden et al. [16], Ueshiba & Tomita [17],

and Mahamud & Hebert [18] proposed iterative methods to

solve the Sturm/Triggs projective factorization problem.

A ﬁnal class of methods uses the perspective camera

model, i.e., the intrinsic parameters are known: this model

is described by eqs. (1) and (2) or, equivalently by eq. (6). If

the perspective depths λ

ij

are known, the perspective factor-

ization problem is identical to the afﬁne factorization prob-

lem. Christy & Horaud [19] and Fujiki & Kurata [20] sug-

gested a method that consists in incrementally estimating

the depth parameters starting with a weak-perspective (or a

paraperspective) approximation (λ

ij

= 1). Each iteration

updates the perspective depths and allows to solve for the

reversal ambiguity. At the limit, these algorithms converge

to an accurate perspective solution. There has been no at-

tempt to incorporate robustness in such iterative perspective

factorization procedures.

2. Robust perspective factorization

The perspective factorization method that we propose builds

on the work of [5], [19], and [13] brieﬂy described above.

Our algorithm estimates values for the perspective depth it-

eratively in a robust fashion starting from a weak perspec-

tive camera model. The robustness is deﬁned two-fold: with

respect to the a-priori knowledge of both noise covariance

and missing data, and with respect to the adaptive knowl-

edge of the residual between weak-perspective and perspec-

tive image measurements of each of the points. The algo-

rithm requires S, W and the intrinsic camera parameters as

input and provides the extrinsic camera parameters and the

3-D coordinates of the points as output. Since the proposed

algorithm puts emphasis on the camera calibration (thus the

M), it can dispense of the 2-D points that are far from the

center of projection and thus provide a bad initial estimate

for weak perspective approximation.

An outline of the algorithm is provided below:

1. ∀i, i ∈ {1...k} and ∀j, j ∈ {1...n} set: ε

ij

= 0;

2. Update the values of s

ij

using ε

ij

as shown in (5);

3. Perform an Euclidean reconstruction with a weak per-

spective camera:

a) Perform a robust afﬁne reconstruction up to a

afﬁne transformation using PowerFactorization

(given S and W).

b) Perform an Euclidean reconstruction by impos-

ing orthonormality constraints on the 3×3 matrix

T.

4. ∀i, i ∈ {1...k} and ∀j, j ∈ {1...n} estimate new

values for ε

ij

using(3);

3

5. Re-weight the values of w

ij

corresponding to non-

missing data using a robust error function and the esti-

mate ε

ij

s

ij

.

6. Check the values of ε

ij

:

if ∀(i, j) the values of ε

ij

just estimated at this it-

eration are identical with the values estimated at

the previous iteration,

then stop;

else go to step 2.

P

o

P

i

optical axis

image plane

first iteration

second iteration

...

last iteration

center of

projection

not modified

weak perspective

projection

perspective

projection

Figure 2: The iterative algorithm modiﬁes the projection

of a 3-D point from true perspective to weak perspective.

Taken from [19]

A visual representation of the iterations of the algo-

rithm can be observed in ﬁgure 2. Step 1 initializes ε

ij

=

0, ∀(i, j), which is the same as approximating the camera

with the weak perspective model. Step 3 adapts the weights

w

ij

based on the 2-D reprojection error e

ij

obtained be-

tween the perspective model and the weak-perspective cam-

era model.

e

ij

= ||((1+ε

ij

)s

ij

−s

0

i

)−(s

ij

−s

0

i

)||

2

= ||ε

ij

s

ij

||

2

(11)

The entries w

ij

are reweighted using the truncated quadratic

as the robust error function

1

:

w

ij

(e

ij

) =

1 e

ij

< h

h

2

e

2

ij

e

ij

> h

(12)

At initialization, there is no knowledge about the per-

spective depths, therefore e

ij

= 0 for all the points and

all the cameras. Once an afﬁne reconstruction is estimated,

one can measure the quality of the weak-perspective model

through eq. (11) and incrementally reweight the measure-

ments in order to favour points that best satisfy the weak-

perspective model.

1

In all our experiments we set h = 0.02.

Step 4 is broken into two stages. The afﬁne reconstruc-

tion is recovered using PowerFactorization as the robust

afﬁne factorization method of choice (other methods have

also been considered - see section 3.1) As detailed earlier,

such a reconstruction is obtained only up to an afﬁne trans-

formation (T). The Euclidean reconstruction is obtained by

imposing orthonormality constraints on the 3 ×3 matrix T.

We have adopted the elegant solution proposed by [13] to

recover the T matrix.

A reversal ambiguity is inherent to any factorization

based technique, since (7) can be written as S = MP =

(−M)(−P). In other words,

ε

ij

= ±

r

z

i

· P

j

t

z

i

(13)

Therefore, two values for ε

ij

are estimated at each iteration

of the algorithm. After N iterations however, there will be

2

N

possible solutions. Fortunately, not all the solutions are

consistent with the image data and a veriﬁcation technique

(the smallest cumulative 2-D reprojection error) allows to

check this consistency and to avoid the combinatorial ex-

plosion of the number of solutions.

3. Multiple-camera calibration

In this section we describe how the solution obtained in

the previous section is used within the context of multiple-

camera calibration. As already described above, we are

interested in the estimation of the external camera param-

eters, i.e., the alignment between a global reference frame

(or the calibration frame) and the reference frame associated

with each one of the cameras. We assume that the internal

camera parameters were accurately estimated using avail-

able software. There may be many cameras. In practice we

tested our algorithm with two setups, each composed of 30

cameras. Finding point correspondences accross the images

provided by such a setup is an issue in its own right because

one has to solve for a multiple wide-baseline point corre-

spondence problem. We will brieﬂy describe the practical

solution that we retained and which avoids badly matched

points. The Euclidean shape and motion solution found in

section 2 must be aligned with the global reference frame

associated with the multiple-camera setup, i.e., the calibra-

tion frame: rotation, translation and scale. We will explain

how this alignment can be robustly estimated.

Multiple-camera data acquisition. In order to perform

multiple-camera acquisition and multiple-image matching,

we use a simple linear object composed of four identical

markers with known one-dimensional coordinates. Using

the one-dimensional coordinates of the markers and their

4

cross-ratio (a projective invariant) we obtain 3-D to 2-D as-

signments between the markers and their observed image

locations. From these assignments one can easily obtain

2-D to 2-D matches over the camera frames, at some time

t. The cameras are ﬁnely synchronized, therefore the lin-

ear object can be freely moved through the 3-D space and

cover the common ﬁeld of view of the cameras. In prac-

tice we accumulate data over several frames. In the two

examples below we used 58 frames and 74 frames, i.e., 232

points and 296 points, respectively. Figure 3 depicts several

situations. The top row shows the ideal case where the four

markers can be easily identiﬁed in the images. The middle

row shows degenerate views due to perspective foreshort-

ening. The bottom row shows a situation where markers are

missing. Uncertain 2-D locations as well as missing data

are described by an a priori weight matrix W.

Figure 3: Examples of calibration input data: (a) typical

frames; (b) frames where due to the perspective distortion

the view is degenerate; (c) frames where some of the control

points are missing due to the limited ﬁeld of view of the

camera

Alignment. We denote by r

i

and t

i

(the rotation matrix

and translation vector) the parameters of camera i found by

factorization. The 3×4 projection matrix associated with

camera i writes:

r

i

t

i

(14)

We also denote by r, t, and s the alignment parameters – ro-

tation, translation, and scale – allowing to map the calibra-

tion frame onto the factorization frame. We also consider

m control points, m << n, i.e., the total number of points.

Let P

j

be the coordinates of the control points in the factor-

ization frame and Q

j

be their coordinates in the calibration

frame. The optimal alignment parameters may be found by

minimizing:

min

s,r,t

m

¸

j=1

srQ

j

+t −P

j

2

(15)

The minimizer of this error function can be found in closed

form using unit quaternions [21] to represent the rotation r

or with dual-number quaternions [22] to represent the rigid

displacement. Therefore, the new projection matrix asso-

ciated with camera i, that aligns the calibration frame with

the image-plane coordinates, writes as:

r

i

t

i

¸

sr t

0

1

r

i

ti

s

¸

r

t

s

0

1

(16)

3.1. Simulated data

In all our experiments we used two camera setups. Both

setups use 30 identical cameras whose intrinsic parameters

were estimated in advance. We also estimated their exter-

nal parameters using a non-linear method based on bundle

adjustment. The external parameters obtained by bundle

adjustment are the ground truth. The ﬁrst setup is shown

in ﬁgure 8 and is referred to as the line case. The second

setup is illustrated in ﬁgure 1 and is referred to as the cor-

ner case). We simulated various conﬁgurations of 200 3-D

data points which are projected onto the 30 images, using

the camera parameters provided by bundle adjustment (the

ground truth).

The simulated experiments compare the results of our

method (robust perspective factorization) with the weak-

perspective factorization methods propsed by Tomasi &

Kanade [1] – the SVD method, Hartley & Schaffalitzky [5]

– PowerFactorization, Aanaes et al. [13] – adaptive weight-

ing, and Buchanan & Fitzgibbon [14] – damped Newton.

Through these experiments we measured the RMS of the 2-

D reprojection errors when increasing noise is added to the

data and when the number of missing entries in the mea-

surement matrix increases as well.

Figures 4 and 5 summarize the results of these simula-

tions with the ﬁrst camera setup (the line case) and ﬁgures 6

and 7 summarize the results obtained with the second cam-

era setup (the corner case).

The best and most consistent results are obtained with

afﬁne PowerFactorization and with our method (perspective

factorization that uses PowerFactorization in its inner loop).

It is interesting to notice that with increasing noise, the per-

formance of these two methods are comparable. This is due

to the fact that when high-amplitude noise is present in the

data, the incremental perspective algorithmcannot currently

5

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

10

0

10

1

10

2

10

3

10

4

% of Data Missing

R

M

S

P

i

x

e

l

E

r

r

o

r

affine−svd

affine−powerfactorization

affine−aanaes

affine−damped newton

persp.−powerfactorization

Figure 4: Line Case - The algorithm described in section 2

is compared with different afﬁne factorization techniques:

robustness towards missing features. The 2-D reprojection

error is measured in in pixels.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

10

0

10

1

10

2

10

3

10

4

% of Data Noised with sigma=0.5

R

M

S

P

i

x

e

l

E

r

r

o

r

affine−svd

affine−powerfactorization

affine−aanaes

affine−damped newton

persp.−powerfactorization

Figure 5: Line Case - The algorithm described in section 2

is compared with different afﬁne factorization techniques:

robustness towards noise in measurements.

distinguish between projection errors and errors to to noisy

measurements.

eq

3.2. Experimental data

In this subsection we present some of the calibration results

obtained with the two camera setups and using the exper-

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

10

0

10

1

10

2

10

3

10

4

% of Data Missing

R

M

S

P

i

x

e

l

E

r

r

o

r

affine−svd

affine−powerfactorization

affine−aanaes

affine−damped newton

persp.−powerfactorization

Figure 6: Corner Case - The algorithm described in section

2 is compared with different afﬁne factorization techniques:

robustness towards missing features.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

10

1

10

2

10

3

10

4

% of Data Noised with sigma=0.5

R

M

S

P

i

x

e

l

E

r

r

o

r

affine−svd

affine−powerfactorization

affine−aanaes

affine−damped newton

persp.−powerfactorization

Figure 7: Corner Case - The algorithm described in section

2 is compared with different afﬁne factorization techniques:

robustness towards noise in measurements. The 2-D repro-

jection error is measured in in pixels.

imental data. The performance obtained with the method

described in section 2 is compared to both weak-perspective

algorithms and to the ground-truth.

The corner case uses 296 points accumulated over

74 frames. Therefore the measurement matrix has

2×30×296 = 17760 entries, out of which 37% are miss-

ing. The recovered camera layout is shown in ﬁgure 1 and

6

a comparison with the ground-truth is detailed in table 1.

The translation error is represented as percentage error with

respect to the groundtruth estimate, whereas the rotation

error is represented in degrees. For completeness, table 2

shows the mean errors when applying the iterative perspec-

tive algorithm in conjunction with various afﬁne factoriza-

tion methods. This table also compares the results with the

results obtained through purely weak-perspective methods.

The 3-D point error is measured in milimeters, whereas the

2-D error is measured in pixels.

Camera Rot. Tr. Camera Rot. Tr.

1 3.31 19.47 16 3.82 18.61

2 4.11 16.89 17 2.44 18.44

3 5.31 16.39 18 0.78 19.23

4 4.31 16.58 19 2.41 22.63

5 4.24 17.52 20 4.27 28.35

6 5.04 16.44 21 2.54 22.44

7 4.75 17.90 22 3.76 21.50

8 4.15 18.13 23 2.88 19.91

9 4.25 17.72 24 1.72 23.07

10 5.19 16.85 25 0.65 22.81

11 3.82 17.84 26 4.07 21.12

12 4.07 18.03 27 5.23 27.09

13 4.53 18.32 28 4.66 22.83

14 2.23 19.84 29 2.06 24.10

15 2.92 18.71 30 5.22 26.64

Table 1: Corner Case - Rotation and translation errors ob-

tained for each individual camera. Rotation is measured in

degrees, and the translation is measured in percentage error

with respect to the original groundtruth

Method 2-D 3-D Cal.Rot Cal.Tr

A

f

ﬁ

n

e

SVD 156 1405 11 48

PowerFact 16 174 4 14

Aanaes - - - -

Damped Newton 892 325 77 119

P

e

r

s

p

.

SVD 82 1558 56 41

PowerFact 13 285 4 20

Aanaes - - - -

Damped Newton 480 793 10 90

Table 2: Comparison between afﬁne and perspective for

corner case

The line case uses 232 points accumulated over 58

frames and there is no missing data in this case. The re-

sulting camera layout is shown in ﬁgure 9. This result may

be compared with the ground-truth shown in ﬁgure 8. Table

3 summarizes the comparison of our method with various

weak-perspective methods. Again, the ﬁgures indicated the

discrepencies with respect to the ground-truth. As it can be

observed in table 3, the perspective reconstruction provides

clear beneﬁts over the weak perspective model. The effect

is diminished when the weak perspective approximations is

close to the perspective camera model, as shown in table 2.

Method 2-D 3-D Cal.Rot Cal.Tr

A

f

ﬁ

n

e

SVD 191 1058 80 58

PowerFact 37 307 9 42

Aanaes 191 1058 80 58

Damped Newton 892 325 77 119

P

e

r

s

p

.

SVD 125 1025 70 76

PowerFact 3 164 3 5

Aanaes 124 1031 77 75

Damped Newton 255 357 50 127

Table 3: Comparison between afﬁne and perspective for line

case

Figure 8: Line Case - Groundtruth Camera Layout (232 fea-

tures - no missing data)

Figure 9: Line Case - Reconstructed Camera Layout (232

features - no missing data)

4. Discussion

In order to interpret these results, it is worth mentioning

the experimental conditions. In the line case 3-D points

are situated at approximatively 1.8 meters away from the

cameras. In the corner case the 3-D points are at 5.1 me-

7

ters away from the cameras. Therefore the second camera

setup is consistent with both the weak-perspective and the

perspective models. The advantage of perspective factor-

ization is clearly shown with the ﬁrst example. It is worth-

while to mention that in the line case the cameras’ optical

axes roughly lie in the same plane and they have a common

rotation axis, and therefore the motion matrices associated

with the extrinsic camera parameters are linearly dependent,

which makes the M matrix a rank-2 matrix. Factorization

methods assume the M matrix to be a rank-3 matrix.

We plan to statistically characterize the perspective error

and to include such prior knowledge in the iterative perspec-

tive factorization algorithm.

Acknowledgments. This research has been supported

by the European Project VISIONTRAIN.

References

[1] C. Tomasi and T. Kanade, “Shape and motion from image

streams under orthography: A factorization mehod,” Inter-

national Journal of Computer Vision, vol. 9, no. 2, pp. 137–

154, Nov. 1992.

[2] R. Hartley and A. Zisserman, Multiple View Geometry in

Computer Vision, Cambridge University Press, Cambridge,

UK, 2000.

[3] Bill Triggs, Philip McLauchlan, Richard Hartley, and An-

drew Fitzgibbon, Bundle Ajustment — A Modern Synthesis,

vol. 1883 of Vision Algorithms: Theory and Practice, pp.

298–372, Spriner, lncs edition, 2000.

[4] P. Anandan and M.Irani, “Factorization with uncertainty,”

International Journal of Computer Vision, vol. 49, no. 2/3,

pp. 101–116, 2002.

[5] R. Hartley and F. Schaffalitzky, “Powerfactorization: 3d re-

construction with missing or uncertain data,” Tech. Rep.,

Australian National University, 2003.

[6] R. Hartley D. Q. Huynh and A. Heyden, “Outlier correction

in image sequences for the afﬁne camera,” in Proceedings

ICCV, 2003, vol. 1, pp. 585–590.

[7] R. Vidal and R. Hartley, “Motion segmentation with missing

data using powerfactorization and gpca.,” in Proceedings

CVPR, 2004, vol. 2, pp. 310–316.

[8] K.Ikeuchi H. Shum and R. Reddy, “Principal component

analysis with missing data and its application to polyhedral

object modeling,” IEEE Transactions on Pattern Analysis

and Machine Intelligence, vol. 17, no. 9, pp. 855–867, 1995.

[9] S. Brant, “Closed-form solutions for afﬁne reconstruction

under missing data,” in Proceedings of Statistical Methods

for Video Processing (ECCV ’02 Workshop), 2002, pp. 109–

114.

[10] T. Wiberg, “Computation of principal components when data

are missing,” in Proceedings Symposium of Computational

Statistics, 1976, pp. 229–326.

[11] A. Gruber and Y. Weiss, “Factorization with uncertainty and

missing data: exploring temporal coherence,” in Proceedings

Neural Information Processing Systems (NIPS’2003), 2003.

[12] A. Gruber and Y. Weiss, “Multibody factorization with un-

certainty and missing data using the em algorithm,” in Pro-

ceedings Conference on Computer Vision and Pattern Recog-

nition, 2004, pp. 707–714.

[13] R. Fisker H. Aanaes and K. Astrom, “Robust factorization,”

IEEE Transactions of Pattern Analysis and Machine Intelli-

gence, vol. 24, no. 9, pp. 1215–1225, Sept. 2002.

[14] A. M. Buchanan and A. W. Fitzgibbon, “Damped newton

algorithms for matrix factorization with missing data,” in

CVPR05, 2005, vol. 2, pp. 316–322.

[15] Peter Sturm and Bill Triggs, “A factorization based algo-

rithm for multi-image projective structure and motion,” in

Proceedings of the 4th European Conference on Computer

Vision, Cambridge, England, April 1996, vol. 1065 of LNCS,

pp. 709–720.

[16] A. Heyden R. Berthilsson and G. Sparr, “An iterative fac-

torization method for projective structure and motion from

image sequences,” Image and Vision Computing, vol. 17,

no. 13, pp. 981–991, November 1999.

[17] T. Ueshiba and F. Tomita, “A factorization method for pro-

jective and euclidean reconstruction from multiple perspec-

tive views via iterative depth estimation,” in Proceedings of

Fifth European Conference of Computer Vision, 1998, pp.

296–310.

[18] Shyjan Mahamud and Martial Hebert, “Iterative projective

reconstruction from multiple views,” in Proceedings of the

IEEE Conference on Computer Vision and Pattern Recogni-

tion (CVPR ’00), June 2000, vol. 2, pp. 430 – 437.

[19] S. Christy and R. Horaud, “Euclidean shape and motion from

multiple perspective views by afﬁne iterations,” IEEE Trans-

actions on Pattern Analysis and Machine Intelligence, vol.

18, no. 11, pp. 1098–1104, November 1996.

[20] J. Fujiki and T. Kurata, “Recursive factorization method for

the paraperspective model based on the perspective projec-

tion,” in Proceedings of the International Conference on

Pattern Recognition, 2000, pp. 406–410.

[21] B.K.P. Horn, “Closed-form solution of absolute orientation

using unit quaternions,” J. Opt. Soc. Amer. A., vol. 4, no. 4,

pp. 629–642, 1987.

[22] M. W. Walker, L. Shao, and R. A. Volz, “Estimating 3-d lo-

cation parameters using dual number quaternions,” CGVIP-

Image Understanding, vol. 54, no. 3, pp. 358–367, Novem-

ber 1991.

8

Previous work There are several ways for solving the afﬁne factorization problem. i ∈ {1 .(a) Partial view of a 30-camera setup (b) Calibration data gathered by moving a stick (c) External multi-camera calibration Figure 1: Outline of the multiple-camera calibration method described in this paper. More generally. the singular value decomposition of the measurement matrix provides an optimal solution. The Frobenius norm becomes in this case: M. Initially it was proposed to minimize the Frobenius norm by Tomasi & Kanade [1]: min S − MP 2 F 0 R y · Pj + y i ry · Pj + ty i i = i z · P + tz ri 1 + εij j i (8) where a 3-D point Pj is projected onto the i-th camera at location sij . under weak perspective λij = 1. obtained with 30 cameras in two different layouts. ∀(i. The i following notations have been introduced: Rx = rx /tz . Dividing the above equations with the depth tz we obtain a similar set of scaled equations. x0 = tx /tz and yi = ty /tz are i i i i i i i i the camera coordinates of a point s0 which is the projection i of P0 (the origin of the 3-D frame). j) and we have: S = MP (7) 1. They use the Mahalanobis norm instead of the Frobenius norm. .e. Moreover.1.P The perspective projection equations (1) and (2) can now be rewritten as: λij sij − s0 = i Rx i Rx i Pj (5) min W ⊗ (S − MP) 2 F (9) For k cameras and n points. They reformulate the factorization problem such that the Mahalanobis norm can be transformed into a Frobenius norm. k}. . j ∈ {1 . the model above becomes simpler. eq. (7). This sign ambiguity corresponds to the well known reversal ambiguity when the 3-D world is observed with afﬁne cameras. a central idea is to introduce a weight matrix W of the same size as the measurement matrix S. that can be determined using Euclidean constraints. .. The rotation matrix ri = [rx ry rz ] and the i i i translation vector ti = (tx ty tz ) correspond to the external i i i camera parameters. So far we considered a perspective camera model. Also. The latter is deﬁned up to an afﬁne transformation T. i i i 0 Ry = ry /tz . n} one can write: Sλ = MP (6) 2 where ⊗ is the Hadamard product: A = B ⊗ C ⇐⇒ aij = bij cij . We analyse the performance of our method in the presence of noise and of missing data. . Anandan & Irani [4] extended the classical SVD approach to deal with the case of directional uncertainty. It is worthwhile to notice that both T and −T can be used. In particular. εij denotes the following ratio: rz · Pj (3) εij = i z ti We also introduce the perspective depth λ: λij = 1 + εij (4) In the absence of noise.P 1. If an afﬁne camera model is used instead. The weights wij describe the conﬁdence that .2. i. Finally they develop a modiﬁed factorization technique. MP = MTT−1 P. Problem formulation We consider a calibrated perspective camera model in which case the relationship between 3-D points and image points expressed in camera coordinates can be written as: xij = yij = rx i rz i · Pj + · Pj + tx i tz i = Rx i · Pj + 1 + εij x0 i (1) (2) M.

. If the perspective depths λij are known. .. Sturm and Triggs [15] introduced the use of factorization in multiple image projective structure from motion. 4. They review the alternation methods and they compare their speed of convergence with a damped Newton method (a variation of the Levenberg-Marquardt method). it requires a lot less computation than SVD (5% of the time for a 500 × 500 matrix) and it deals well with missing data (60-90%). A ﬁnal class of methods uses the perspective camera model. The most accurate results are obtained through bundle adjustment methods [2]. The most common way of minimizing eq. At the limit. Another way to alternate between motion and shape estimation is to use the EM algorithm and a nice formulation that uses factor analysis is provided by Gruber and Weiss et al. An outline of the algorithm is provided below: 1. using an Mestimator. The PowerFactorization method [5]. i ∈ {1..we have in an image datum and the values of these weights vary from 1 (highly conﬁdent data) to 0 (missing data).. it can dispense of the 2-D points that are far from the center of projection and thus provide a bad initial estimate for weak perspective approximation. [9].. (9). and ﬁnally the image residuals are calculated and the data (the weights) are modiﬁed. PowerFactorization [5] presents a number of attractive features: It converges by starting with any random input.. ∀i. [12]. (9) with the original data: M.. [6]. The method proposed in [13] uses eq. (8) with modiﬁed data S is the same as eq. The algorithm requires S. Perform an Euclidean reconstruction with a weak perspective camera: a) Perform a robust afﬁne reconstruction up to a afﬁne transformation using PowerFactorization (given S and W). the intrinsic parameters are known: this model is described by eqs. 2.e. Christy & Horaud [19] and Fujiki & Kurata [20] suggested a method that consists in incrementally estimating the depth parameters starting with a weak-perspective (or a paraperspective) approximation (λij = 1). The approximation is performed by modifying the original ˜ data S such that the solution to eq. The robustness is deﬁned two-fold: with respect to the a-priori knowledge of both noise covariance and missing data. and with respect to the adaptive knowledge of the residual between weak-perspective and perspective image measurements of each of the points. Update the values of sij using εij as shown in (5). getting a temporary optimum. i ∈ {1. Our algorithm estimates values for the perspective depth iteratively in a robust fashion starting from a weak perspective camera model.P 2 F (10) The weights are updated at each iteration. (9) is to use alternation methods: these methods are based on the observation that if either M or P is known. This method may well be viewed as both an iterative and an alternation method because the matrix M is estimated using SVD.k} and ∀j. Their formulation is similar in spirit to eq. There has been no attempt to incorporate robustness in such iterative perspective factorization procedures. Each iteration updates the perspective depths and allows to solve for the reversal ambiguity. (9) through non-linear minimization over the motion and shape parameters. j ∈ {1. [16]. Since the proposed algorithm puts emphasis on the camera calibration (thus the M). The latters are not well suited for accurate estimation of motion and shape parameters. [11]. The factorization methods mentioned so far can only deal with afﬁne camera models. 2.. ∀i. b) Perform an Euclidean reconstruction by imposing orthonormality constraints on the 3×3 matrix T. [19].n} estimate new values for εij using(3). robustness both to noisy data and to outliers can be achieved through adaptive weighting by iteratively re-estimating the weight matrix W which amounts to iteratively modify the data S. W and the intrinsic camera parameters as input and provides the extrinsic camera parameters and the 3-D coordinates of the points as output.n} set: εij = 0. Subsequently. (9). (8) in conjunction with a robust weighting matrix (truncated quadratic or Huber’s M-estimator) to iteratively approximate eq. i. (6). [13]. [3] which require proper initialization. [7] as well as many other methods [8]. j ∈ {1. [10] fall into this category.. Alternatively. Buchanan & Fitzgibbon [14] propose to solve eq. (6) and their method relies on fundamental matrices to solve for the projective depths. They propose a hybrid method that switches from alternation to damped Newton.k} and ∀j. Heyden et al. Ueshiba & Tomita [17]. than the P is estimated knowing M. equivalently by eq. 3. these algorithms converge to an accurate perspective solution.P and Mahamud & Hebert [18] proposed iterative methods to solve the Sturm/Triggs projective factorization problem. It is more difﬁcult to solve for the projective multiplecamera calibration and/or reconstruction problem. there is a closed-form solution for the other that minimizes eq. and [13] brieﬂy described above. (1) and (2) or. 3 min W ⊗ (S − MP) 2 F ˜ = min S − MP M. the perspective factorization problem is identical to the afﬁne factorization problem. Robust perspective factorization The perspective factorization method that we propose builds on the work of [5].

Once an afﬁne reconstruction is estimated.see section 3. therefore eij = 0 for all the points and all the cameras. We will explain how this alignment can be robustly estimated.. We will brieﬂy describe the practical solution that we retained and which avoids badly matched points. one can measure the quality of the weak-perspective model through eq. since (7) can be written as S = MP = (−M)(−P). two values for εij are estimated at each iteration of the algorithm. i.e. else go to step 2. Step 4 is broken into two stages. translation and scale.. In practice we tested our algorithm with two setups. such a reconstruction is obtained only up to an afﬁne transformation (T). In other words. 1 In all our experiments we set h = 0. Using the one-dimensional coordinates of the markers and their 4 A visual representation of the iterations of the algorithm can be observed in ﬁgure 2. i. There may be many cameras.. last iteration perspective projection εij = ± weak perspective projection rz · Pj i tz i (13) Po center of projection not modified optical axis image plane Therefore. . Multiple-camera data acquisition. Pi first iteration second iteration . 6.5. which is the same as approximating the camera with the weak perspective model. (11) and incrementally reweight the measurements in order to favour points that best satisfy the weakperspective model. The Euclidean shape and motion solution found in section 2 must be aligned with the global reference frame associated with the multiple-camera setup. After N iterations however. the alignment between a global reference frame (or the calibration frame) and the reference frame associated with each one of the cameras. eij = ||((1+εij )sij −s0 )−(sij −s0 )||2 = ||εij sij ||2 (11) i i The entries wij are reweighted using the truncated quadratic as the robust error function1 : 1 wij (eij ) = h2 e2 ij eij < h eij > h (12) At initialization. j).e. each composed of 30 cameras. A reversal ambiguity is inherent to any factorization based technique. The afﬁne reconstruction is recovered using PowerFactorization as the robust afﬁne factorization method of choice (other methods have also been considered . Step 3 adapts the weights wij based on the 2-D reprojection error eij obtained between the perspective model and the weak-perspective camera model. Figure 2: The iterative algorithm modiﬁes the projection of a 3-D point from true perspective to weak perspective. Check the values of εij : if ∀(i. We have adopted the elegant solution proposed by [13] to recover the T matrix. we use a simple linear object composed of four identical markers with known one-dimensional coordinates. we are interested in the estimation of the external camera parameters. We assume that the internal camera parameters were accurately estimated using available software. As already described above.02. Multiple-camera calibration In this section we describe how the solution obtained in the previous section is used within the context of multiplecamera calibration. The Euclidean reconstruction is obtained by imposing orthonormality constraints on the 3 × 3 matrix T. In order to perform multiple-camera acquisition and multiple-image matching. there will be 2N possible solutions. ∀(i. there is no knowledge about the perspective depths. Taken from [19] 3. Step 1 initializes εij = 0. Re-weight the values of wij corresponding to nonmissing data using a robust error function and the estimate εij sij .. j) the values of εij just estimated at this iteration are identical with the values estimated at the previous iteration.1) As detailed earlier. not all the solutions are consistent with the image data and a veriﬁcation technique (the smallest cumulative 2-D reprojection error) allows to check this consistency and to avoid the combinatorial explosion of the number of solutions. Finding point correspondences accross the images provided by such a setup is an issue in its own right because one has to solve for a multiple wide-baseline point correspondence problem. Fortunately. the calibration frame: rotation. then stop.

We denote by ri and ti (the rotation matrix and translation vector) the parameters of camera i found by factorization.e. The external parameters obtained by bundle adjustment are the ground truth. From these assignments one can easily obtain 2-D to 2-D matches over the camera frames. Uncertain 2-D locations as well as missing data are described by an a priori weight matrix W. Therefore. Hartley & Schaffalitzky [5] – PowerFactorization. using the camera parameters provided by bundle adjustment (the ground truth). and scale – allowing to map the calibration frame onto the factorization frame.e. We simulated various conﬁgurations of 200 3-D data points which are projected onto the 30 images. The cameras are ﬁnely synchronized. respectively. (b) frames where due to the perspective distortion the view is degenerate. the incremental perspective algorithm cannot currently 5 Alignment. translation. and Buchanan & Fitzgibbon [14] – damped Newton. 232 points and 296 points. The middle row shows degenerate views due to perspective foreshortening. This is due to the fact that when high-amplitude noise is present in the data. Both setups use 30 identical cameras whose intrinsic parameters were estimated in advance. frame. [13] – adaptive weighting. Simulated data In all our experiments we used two camera setups. i. that aligns the calibration frame with the image-plane coordinates. the total number of points.r.cross-ratio (a projective invariant) we obtain 3-D to 2-D assignments between the markers and their observed image locations. Figure 3: Examples of calibration input data: (a) typical frames. The ﬁrst setup is shown in ﬁgure 8 and is referred to as the line case. i. Figure 3 depicts several situations. The top row shows the ideal case where the four markers can be easily identiﬁed in the images. In the two examples below we used 58 frames and 74 frames. the performance of these two methods are comparable. The bottom row shows a situation where markers are missing. t. m << n. We also estimated their external parameters using a non-linear method based on bundle adjustment. the new projection matrix associated with camera i.t j=1 srQj + t − Pj 2 (15) The minimizer of this error function can be found in closed form using unit quaternions [21] to represent the rotation r or with dual-number quaternions [22] to represent the rigid displacement. The best and most consistent results are obtained with afﬁne PowerFactorization and with our method (perspective factorization that uses PowerFactorization in its inner loop). It is interesting to notice that with increasing noise. The optimal alignment parameters may be found by minimizing: m min s. therefore the linear object can be freely moved through the 3-D space and cover the common ﬁeld of view of the cameras. We also consider m control points. at some time t. Through these experiments we measured the RMS of the 2D reprojection errors when increasing noise is added to the data and when the number of missing entries in the measurement matrix increases as well. (c) frames where some of the control points are missing due to the limited ﬁeld of view of the camera The simulated experiments compare the results of our method (robust perspective factorization) with the weakperspective factorization methods propsed by Tomasi & Kanade [1] – the SVD method. Aanaes et al.. writes as: ri ti sr 0 t 1 ri ti s r 0 t s 1 (16) 3. The second setup is illustrated in ﬁgure 1 and is referred to as the corner case).. In practice we accumulate data over several frames. and s the alignment parameters – rotation. Let Pj be the coordinates of the control points in the factorization frame and Qj be their coordinates in the calibration .1. The 3×4 projection matrix associated with camera i writes: r i ti (14) We also denote by r. Figures 4 and 5 summarize the results of these simulations with the ﬁrst camera setup (the line case) and ﬁgures 6 and 7 summarize the results obtained with the second camera setup (the corner case).

10 4 affine−svd affine−powerfactorization affine−aanaes affine−damped newton persp.5 0.8 0.9 1 Figure 5: Line Case . imental data.−powerfactorization 10 RMS Pixel Error 10 3 RMS Pixel Error 10 2 2 10 1 10 10 0 1 0 0.6 % of Data Missing 0.2 0.−powerfactorization 10 3 RMS Pixel Error 10 2 RMS Pixel Error 10 2 10 1 10 1 10 0 10 0 0.5 % of Data Missing 0.5 0.4 0.8 0. out of which 37% are missing. The corner case uses 296 points accumulated over 74 frames.1 0.−powerfactorization 10 3 10 4 affine−svd affine−powerfactorization affine−aanaes affine−damped newton persp.6 0.9 0 0 0. eq Figure 7: Corner Case .1 0. Experimental data In this subsection we present some of the calibration results obtained with the two camera setups and using the exper- .8 0.6 0.The algorithm described in section 2 is compared with different afﬁne factorization techniques: robustness towards noise in measurements.4 0.3 0.−powerfactorization 10 3 affine−svd affine−powerfactorization affine−aanaes affine−damped newton persp.4 0. The recovered camera layout is shown in ﬁgure 1 and 6 3.7 % of Data Noised with sigma=0.8 0.2 0.2 0. The 2-D reprojection error is measured in in pixels.10 4 affine−svd affine−powerfactorization affine−aanaes affine−damped newton persp.The algorithm described in section 2 is compared with different afﬁne factorization techniques: robustness towards missing features.7 0.5 0. The performance obtained with the method described in section 2 is compared to both weak-perspective algorithms and to the ground-truth. Therefore the measurement matrix has 2×30×296 = 17760 entries.The algorithm described in section 2 is compared with different afﬁne factorization techniques: robustness towards missing features. distinguish between projection errors and errors to to noisy measurements.2. The 2-D reprojection error is measured in in pixels.7 0.4 0.5 0.6 0. 10 4 Figure 6: Corner Case .9 1 0 0.3 0.3 0.7 % of Data Noised with sigma=0.The algorithm described in section 2 is compared with different afﬁne factorization techniques: robustness towards noise in measurements.9 1 Figure 4: Line Case .1 0.5 0.2 0.3 0.1 0.

31 4.09 18. Discussion In order to interpret these results.50 18.1 me- Persp.61 16.66 22.Rot Cal.78 19.Tr 156 1405 11 48 16 174 4 14 892 325 77 119 82 1558 56 41 13 285 4 20 480 793 10 90 Figure 8: Line Case .44 16.04 4. and the translation is measured in percentage error with respect to the original groundtruth Method SVD PowerFact Aanaes Damped Newton SVD PowerFact Aanaes Damped Newton Afﬁne 2-D 3-D Cal. The resulting camera layout is shown in ﬁgure 9. the ﬁgures indicated the discrepencies with respect to the ground-truth.39 18 0. table 2 shows the mean errors when applying the iterative perspective algorithm in conjunction with various afﬁne factorization methods.54 22.25 5. This result may be compared with the ground-truth shown in ﬁgure 8. Camera 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 Rot.72 23. This table also compares the results with the results obtained through purely weak-perspective methods.63 17. Table 3 summarizes the comparison of our method with various weak-perspective methods.07 16. Method SVD PowerFact Aanaes Damped Newton SVD PowerFact Aanaes Damped Newton Afﬁne 2-D 191 37 191 892 125 3 124 255 3-D Cal.24 5.83 19.23 16.91 17.31 4.Groundtruth Camera Layout (232 features .81 17.Rotation and translation errors obtained for each individual camera.75 4.no missing data) 4.84 26 4.41 22. Table 2: Comparison between afﬁne and perspective for corner case The line case uses 232 points accumulated over 58 frames and there is no missing data in this case. it is worth mentioning the experimental conditions.11 5.15 4. As it can be 7 Figure 9: Line Case .10 18.23 27. as shown in table 2.89 17 2.32 28 4.71 30 5.06 24.13 23 2. Tr.07 4.23 2.no missing data) Persp. the perspective reconstruction provides clear beneﬁts over the weak perspective model.76 21. The 3-D point error is measured in milimeters.Rot Cal.84 29 2.88 19. In the corner case the 3-D points are at 5.22 26.90 22 3. Camera Rot.19 3. For completeness. 3. .53 2. whereas the rotation error is represented in degrees.Tr 1058 80 58 307 9 42 1058 80 58 325 77 119 1025 70 76 164 3 5 1031 77 75 357 50 127 Table 3: Comparison between afﬁne and perspective for line case Table 1: Corner Case . The translation error is represented as percentage error with respect to the groundtruth estimate.44 17.Reconstructed Camera Layout (232 features .58 19 2.52 20 4.82 18.47 16 3.03 27 5.85 25 0. Again.92 Tr.12 18.31 4. The effect is diminished when the weak perspective approximations is close to the perspective camera model.27 28.65 22.8 meters away from the cameras.44 18. 19.44 21 2.35 16. In the line case 3-D points are situated at approximatively 1.a comparison with the ground-truth is detailed in table 1.72 24 1. Rotation is measured in degrees. whereas the 2-D error is measured in pixels.64 observed in table 3.07 21.82 4.

Ueshiba and F. Wiberg. Heyden R. “Shape and motion from image streams under orthography: A factorization mehod. Brant. “Euclidean shape and motion from multiple perspective views by afﬁne iterations. Fujiki and T. [17] T. no. “A factorization based algorithm for multi-image projective structure and motion. pp. 9. vol. vol. no. “Estimating 3-d location parameters using dual number quaternions.” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR ’00). pp. 296–310. vol. Gruber and Y. “Recursive factorization method for the paraperspective model based on the perspective projection. pp. References [1] C.” in CVPR05. no.ters away from the cameras.” IEEE Transactions of Pattern Analysis and Machine Intelligence. Cambridge. 4. pp.Irani. pp. 2000.” International Journal of Computer Vision. 109– 114. Horaud. Rep. 1883 of Vision Algorithms: Theory and Practice. 2004. pp. L. Sparr. pp. Therefore the second camera setup is consistent with both the weak-perspective and the perspective models. 298–372. 310–316. pp.. “A factorization method for projective and euclidean reconstruction from multiple perspective views via iterative depth estimation. no. Buchanan and A. vol. Philip McLauchlan.” in Proceedings Conference on Computer Vision and Pattern Recognition. Walker.” Tech.” International Journal of Computer Vision. Sept. 13. 2. vol. pp. 137– 154. 9.. Hartley and A. This research has been supported by the European Project VISIONTRAIN. Heyden. Christy and R. “Computation of principal components when data are missing. 2004. [2] R. Hartley. “Damped newton algorithms for matrix factorization with missing data.” IEEE Transactions on Pattern Analysis and Machine Intelligence. vol. pp. W. [5] R.” in Proceedings Neural Information Processing Systems (NIPS’2003). Hartley and F. 1098–1104. “Closed-form solutions for afﬁne reconstruction under missing data.” IEEE Transactions on Pattern Analysis and Machine Intelligence. Fisker H. which makes the M matrix a rank-2 matrix. “Powerfactorization: 3d reconstruction with missing or uncertain data. [4] P. [6] R.” in Proceedings of the International Conference on Pattern Recognition. Nov.” in Proceedings Symposium of Computational Statistics. Tomasi and T. [8] K. 358–367. 8 . 585–590.” in Proceedings of the 4th European Conference on Computer Vision. November 1999. [19] S. We plan to statistically characterize the perspective error and to include such prior knowledge in the iterative perspective factorization algorithm. vol. England. 2000. 1992. Berthilsson and G. 18. “Multibody factorization with uncertainty and missing data using the em algorithm. 1215–1225. “Closed-form solution of absolute orientation using unit quaternions. vol. 1998. Shum and R. 24. lncs edition. 49. Vidal and R. 2000. Anandan and M. 430 – 437. Richard Hartley. Horn. vol. “An iterative factorization method for projective structure and motion from image sequences. [12] A. Schaffalitzky. 709–720. 9. Gruber and Y. “Factorization with uncertainty. [16] A. 629–642. November 1996. A. 2002. Spriner. 406–410. April 1996.” in Proceedings of Statistical Methods for Video Processing (ECCV ’02 Workshop).K. Reddy. Huynh and A. 2005. W. Weiss. Factorization methods assume the M matrix to be a rank-3 matrix. “Principal component analysis with missing data and its application to polyhedral object modeling. pp. The advantage of perspective factorization is clearly shown with the ﬁrst example. Aanaes and K. 2. 11. 2002. Fitzgibbon. Tomita. [10] T. Acknowledgments. pp. June 2000. vol. A. Zisserman. Multiple View Geometry in Computer Vision. and Andrew Fitzgibbon. “Robust factorization. Opt. vol. vol. [13] R. 1995. Australian National University. 54. Cambridge University Press. [7] R. 1987. 2002. “Iterative projective reconstruction from multiple views. It is worthwhile to mention that in the line case the cameras’ optical axes roughly lie in the same plane and they have a common rotation axis. 17. 1976. Astrom. “Motion segmentation with missing data using powerfactorization and gpca. 1. 2/3. M. vol. Bundle Ajustment — A Modern Synthesis. Shao. no. UK. pp. November 1991. [15] Peter Sturm and Bill Triggs. pp.” Image and Vision Computing. no. Amer. pp. and therefore the motion matrices associated with the extrinsic camera parameters are linearly dependent. [22] M. 1065 of LNCS. Hartley D. pp. 981–991. 2003. 707–714. no. Kanade. Weiss. pp. [9] S. “Factorization with uncertainty and missing data: exploring temporal coherence.” in Proceedings of Fifth European Conference of Computer Vision. 101–116.Ikeuchi H.P. 2003. 4.” CGVIPImage Understanding. [21] B. [14] A. 17. [11] A. [18] Shyjan Mahamud and Martial Hebert. 316–322. vol. Cambridge. and R.” J. 855–867. [20] J. Volz. 2. 229–326. Q.. no. [3] Bill Triggs. Kurata. pp. 3. 2003. 2. “Outlier correction in image sequences for the afﬁne camera. Soc.” in Proceedings CVPR. pp.” in Proceedings ICCV.

- What is Computation
- Numerical Methods Book.pdf
- Discrete Inverse Problem - Insight and Algorithms
- pollefeysIJCV04final
- Suggested+Solutions+MaC2NVC08+Part+II+VG MVG Level+Test
- Fireworks Portfolios
- Summer Maths
- fkv,jvkhxjgckmvkckckc
- Gentle Intro
- Geometry of Interaction and Dynamics of Proof Reduction
- The Unfolding of General Petri Nets
- Saharon Shelah and Mor Doron- Bounded m-ary Patch-Width Are Equivalent for m≥3
- Mesh Generation
- Fried Gut
- Untitled
- Samson Abramsky and Ross Duncan- A categorical quantum logic
- Power Series
- PSsoln1
- Event Structures With Symmetry
- Expansion in Finite Simple Groups of Lie Type
- jh
- AP Calculus Notebook
- Limits
- ERDOS-SMARANDACHE NUMBERS
- Limits and Continuity
- Sequences
- Lag Range
- 02_03

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd