J Math Imaging Vis (2017) 58:189–210
DOI 10.1007/s10851-016-0700-6
Robust Fitting of Ellipsoids by Separating Interior and Exterior
Points During Optimization
Ezequiel López-Rubio1 · Karl Thurnhofer-Hemsi1 · Óscar David de Cózar-Macías2 ·
Elidia Beatriz Blázquez-Parra2 · José Muñoz-Pérez1 · Isidro Ladrón de Guevara-López2
Received: 27 January 2016 / Accepted: 9 December 2016 / Published online: 24 December 2016
© Springer Science+Business Media New York 2016
Abstract Fitting geometric or algebraic surfaces to 3D data
is a pervasive problem in many fields of science and engineer-
ing. In particular, ellipsoids are some of the most employed
features in computer graphics and sensor calibrations. They
are also useful in pattern recognition, computer vision, body
detection and electronic device design. Standard ellipsoid
fitting techniques to solve this problem involve the minimiza-
tion of squared errors. However, most of these procedures are
sensitive to noise. Here, we propose a method based on the
minimization of absolute errors. Although our algorithm is
iterative, an adaptive step size is used to achieve a faster con-
vergence. This leads to a substantial improvement in robust-
ness against outlier data. The proposal is demonstrated with
several computational examples which comprise synthetic
data and real data from a 3D scanner and a stereo camera.
Keywords Least absolute error · Ellipsoid fitting · Outliers ·
Robust estimation · Gradient descent
1 Introduction
In recent years, research concerning computer vision has
developed rapidly. Fitting primitive models such as a
quadratic curve or surface to a given cloud of points is
a fundamental task in digital image analysis, visual pat-
B Ezequiel López-Rubio
ezeqlr@lcc.uma.es
1
Department of Computer Languages and Computer Science,
University of Málaga, Bulevar Louis Pasteur, 35, 29071
Málaga, Spain
2 grated calibration of magnetic gradient sensor systems [21].
Department of Graphic Expression, Design and Projects,
University of Málaga, Calle Dr Ortiz Ramos, 29071 Málaga,
Spain
tern recognition, computer vision and computer graphics
with numerous applications. The 2D case involves fitting
a quadratic curve to a set of sample points on the plane
[8,38,49,65,73]. Also, the ellipticity and circularity of the
set of samples can be measured [48]. The 3D case con-
sists in fitting a quadratic surface to a set of points in space
[56], which has applications to 3D reconstruction [12], pose
estimation [39], restricted stereo correspondence problem
[10] and object recognition [5,57]. Moreover, fitting alge-
braic surface with scatter 3D points has been discussed
widely in [40] and some remarkable work has been done
in [34,35,53,63,66,67].
One of the most used geometric primitives is the ellip-
soid, which has a paramount importance in many fields.
For instance, it is employed in calibration for three-axis
magnetic sensors, as shown in [21], where a mathematical
model is proposed for scalar calibration of a single three-axis
magnetometer and a least squares ellipsoid fitting algorithm
to estimate the detailed error parameters. Three-axis mag-
netometers are used to measure the intensity of the Earth
magnetic field, being widely applied in navigation systems
and aircraft [45]. According to Gietzelt et al. [23], it is known
that calibration algorithms for three-dimensional accelerom-
eters and magnetometers are equivalent to a 3D-ellipsoid
fitting problem. Most calibration algorithms are based on
3D ellipsoid fitting methods, which minimize a cost func-
tion to estimate the value of a set of calibration parameters
that define the sensor model. These algorithms are generally
employed in micromagnetometer calibration for accurate ori-
entation estimation [79], micromagnetometers, together with
inertial sensors and also, to calibrate magnetic angular rate
and gravity (MARG) sensors, Olivares et al. [50] or in inte-
Many commercial electronic devices need to solve 3D
ellipsoid fitting problems. Smartphones, tablets, GPS navi-
123
190
gators [16,41], watches and videogame peripherals comprise
accelerometers and magnetometers that are used to estimate
the orientation of a body in space [15], compute trajectories
[36,52,61], analyze motion [9,24,64], detect obstacle col-
lision [54], and other innovative measurement applications
such as 3D scanning [27] or in reverse engineering by a laser
scanner point data from the surface of an object [43].
In the medical field, fitting ellipsoids has been used for
estimating diameters of pulmonary nodules [37] as a tech-
nique to extract a pulmonary nodule from helical thoracic
computed tomography (CT) scans and estimate its diame-
ter. Another outstanding study is presented by Ge et al. [22],
in which a computer-aided detection system is developed to
assist radiologists in the detection of lung nodules on tho-
racic CT images, using a 3D gradient field method and 3D
ellipsoid fitting. Furthermore, Karnesky et al. [33] has studied
the coalescence of γ  (L12 ) precipitates in Ni–Al–Cr through
the best-fit ellipsoids of atom-probe tomographic data, since
best-fit ellipsoids have equivalent centroids, moments of iner-
tia, and principle axes for arbitrarily shaped precipitates.
Additionally, [44] suggested a 3D semi-automatic prostate
segmentation method for B-mode Trans-Rectal UltraSound
(TRUS) images, where it is assumed that the prostate
transversal section shape may be a tapered ellipse.
Other research fields are focused on body detection or
classification, such as nose tip detection by Sarakon et al.
[58]. This is useful because the nose is a facial feature whose
shape is not affected by facial expressions and is visible for
almost any orientation of the head [68]. Furthermore, the
nose is often used as a baseline to localize other facial land-
marks such as the eyes and the mouth. Other approaches
include [59] which proposes a noncontact method to clas-
sify the face shape by using Support Vector Machine (SVM)
technique and Mahalanobis distance applied to define a pre-
cise head region. Moreover, Grammalidis and Strintzis [26]
suggested a head detection of a person and tracking system
by 2D and 3D ellipsoid fitting by K-means algorithm in head
segmentation and 3D ellipsoid model.
The minimization of the sum of absolute errors (also called
L1-norm minimization) has been used previously to fit sur-
faces in 3D. The method by Flöry and Hofer [19] is aimed
to fit B-spline surfaces, so that the spline control points can
be displaced slightly to deform the surface locally in order
to fit the local features of the point cloud. This requires the
use of a foot point in the surface for each sample point, so
that the distance of the sample to the surface is computed
with respect to the foot point. Then one can use three (non-
squared) distance measures: (a) the distance between the foot
point and the sample point, (b) the distance from the tangent
plane at the foot point to the sample point, (c) a second-order
approximant of the signed distance function, based on the
local Frenet frame associated to the foot point. Depending
on the chosen option the L1 distance minimization is carried
123
J Math Imaging Vis (2017) 58:189–210
out by second-order cone programming, linear programming
or the proximal bundle method. In the linear programming
case, the introduction of auxiliary variables enables the refor-
mulation of the absolute values into linear functions.
Our approach considers a much more restricted class of
surfaces, namely ellipsoids. This means that the surface can-
not be deformed locally, i.e., there are no control points. Foot
points are not necessary in our approach because the ana-
lytic form of the ellipsoid allows computing the deviation
of the sample points from the surface without foot points.
The objective function is nonlinear, so a specifically tuned
version of gradient descent is used instead of linear program-
ming. The role of the reformulation of the absolute value by
auxiliary variables in Flöry and Hofer’s method is accom-
plished in our method by the split of the point cloud into two
sets, namely the interior points and the exterior points.
From the above it can be said that, although the kind of sur-
faces to be fitted and the methodology of our proposal differ
significantly from Flöry and Hofer’s, both take advantage of
the robustness against outliers provided by L1 optimization.
This paper is structured as follows. Section 2 presents
previous works and the definition of earlier ellipsoid fitting
methods. Subsequently, Sect. 3 describes the formulation of
our method and Sect. 4 reports the obtained experimental
results using synthetic and real data. Finally, Sect. 5 summa-
rizes the main conclusions of this research.
2 Previous Work
In the literature, a wide range of fitting techniques have been
applied depending on the researcher. There are two main
categories of fitting methods [45,75], namely using algebraic
distances [2,30] or geometric distances [1,6,20,25,57,62]
according to their error distance definitions. Algebraic fitting
methods often yield a noniterative unique solution, while
geometric ones make use of some nonlinear constraints by
iterative algorithms [49].
In 2004, Li et al. [40] explain that ellipsoid fitting
can be accomplished in many ways by applying some
known bounded surface fitting techniques, see [34] and [67].
Even so, these techniques involve a nonlinear optimization
method, which cannot guarantee an optimal solution and
stops at a local minimum. Their purpose is to develop an
effective and efficient ellipsoid fitting algorithm and not to
consider those fitting methods based on the geometric dis-
tance or the parametric representation of an ellipsoid, since
they give rise to a lengthy nonlinear optimization process.
In contrast, Yu et al. [76] mention that general ellipsoids
do not have a natural geometric definition similar to ellipses
or spheroids. Anyway, they can still generalize the ellipse
fit algorithm to three dimensions in the case of a spheroid,
since a spheroid has the same basic geometric property as an
J Math Imaging Vis (2017) 58:189–210
ellipse. This is the reason that all algorithms proposed in two
dimensions can be easily generalized to three dimensions in
a natural way.
2.1 Algebraic Fitting Methods
According to Malyugina et al. [45], the existing methods
of ellipsoidal fitting are mainly based on the least squares
method. Methods of this group use an algebraic approach
regarding an implicit representation of the quadratic surface
that [40] defines as the locus of points such that their coordi-
nates satisfy the most general equation of the second degree
in three dimensions:
ax 2 + by 2 + cz 2 + 2 f yz + 2gx z + 2hx y
+2 px + 2qy + 2r z + d = 0 (1)
Now, let
I = a+b+c
J = ab + bc + ac − f 2 − g 2 − h 2
 
a h g 
 
K =  h g f 
g f c 
then it is known that they are invariant under rotation and
translation and Eq. (1) represents an ellipsoid if J > 0, I K >
0 in [28].
In [40] it is shown that when a J − I 2 > 0, Eq. (1) must
represent an ellipsoid if a = 4. Hence, some authors that use
the least squares fitting problem based on algebraic distance
must consider this constraint. In Direct method (DM) [40]
the constraint confines the class of ellipsoids to fit to those
whose smallest radius is at least half of the largest radius; in
the Simple method, the ellipsoid is fitted by bisection search
on a in equation a J − I 2 > 0.
Ying et al.’s [75] work intends to find an algebraic method
for multidimensional ellipsoid-specific fitting, rather than fit-
ting general quadratic surfaces. They draw inspiration from
a study on the ellipse-specific fitting approach proposed by
Fitzgibbon et al. [18] since it outperforms other approaches
because it is robust, efficient and easy to implement. How-
ever, they mention that using linear least-square techniques
with a constraint whose degree is greater than 2 is very dif-
ficult to carry out.
Also, as it is well known, a quadratic surface in n-
dimensional space is defined as the locus of the zeros of
a quadratic polynomial. This quadratic polynomial may be
written in a different notation as a real symmetric matrix
of order n + 1 and an (n + 1)-vector. The leading n × n
principal submatrix of the symmetric matrix is positive or
opposite definite whenever the n-D quadratic surface is an
191
n-D ellipsoid. Fortunately, to ensure that a matrix is positive
or opposite definite, the most practical way may be to utilize
semidefinite programming (SDP) as Calafiore proposed [4].
However, in Calafiore’s method, the run time is significantly
long and memory is often exhausted when the number of
fitted points is greater than several thousands. Hence, Ying
et al. [75] proposed a fast algorithm by minimizing a new
defined vector norm of the algebraic residual vector using
SDP by decreasing the dimensions of the SDP problem.
Another approach is Koopmans’s method that generalized
the Least Squared (LS) estimation scheme by developing a
maximum likelihood (ML) estimation. In order to assume
extended observation vectors and LS it computes an algebraic
fit where the norm vector of the ten parameters describing the
ellipsoid algebraically (1) is equal to 1.
2.2 Geometric Fitting Methods
Some methods have a geometrical definition of error dis-
tance, so that the shortest distance from the data point to the
surface is computed. Though those geometric methods give
better solutions, they require more time for evaluation.
In the literature until 2002, Calafiore [4] first proposed
a feasible approach for multidimensional ellipsoid-specific
fitting using semidefinite programming (SDP), which min-
imizes the 2-norm of the algebraic residual vector, who
presented the solution of spherical and ellipsoidal fitting
based on a “difference-of-squares” (DOS) geometric error
criterion for points in n-dimensional space using Eq. (2)
expressed in 3D form:
 2
f (c, A, r ) = d 2 (xi , c) − d 2 (x̃i , c) (2)
where c is the center of ellipsoid, A is a symmetric positive
definite matrix that defines the shape of the ellipsoid, r is
represented by (x − c)T A (x − c), d (a, b) denotes a metric
distance between the points a and b, and x̃i is the point on
the ellipsoid that lies on the ray xi c.
More recent approaches include the Distance method [32]
which intends to calculate the orthogonal distance of the
points projected onto an ellipsoid. Eberly [14] defines the
error as the normal distance from the point of the cloud to
the ellipsoid. The closest point on the ellipsoid is found by a
Newton’s iteration scheme. He develops methods of hyper-
planar fitting and surface fitting for several kinds of quadric
curves and surfaces. In the problem of ellipsoidal fitting he
focuses on axis-aligned ellipsoids.
3 Methodology
Our ellipsoid fitting methodology is presented next. In gen-
eral terms, it is better to consider criteria based on the sum
123
192 J Math Imaging Vis (2017) 58:189–210

of the absolute error values to detect the presence of outliers
[31,47,55,69]. In this paper, we consider a criterion based on
the mean absolute deviation from the level set which defines
the ellipsoid surface. The proposed algorithm provides good
results for a set of points that are not truly geometrically pre-
cise ellipsoids but rough ellipsoid or egg-like shapes, e.g., real
image applications for agricultural images and many other
nearly round shapes.
We start with a review of some basic facts about ellipsoids
(Sect. 3.1). Then the proposed fitting method and its conver-
gence control procedure are stated in Sects. 3.2 and 3.3. After
that, the initialization procedure is detailed in Sect. 3.4 and
a summary of the algorithm is provided in Sect. 3.5. Finally,
a discussion of the most important features of our proposal
is given in Sect. 3.6.
3.1 Preliminaries
The equation of an ellipsoid reads as follows:
Then we substitute (8) into (5), which leads to (3) by appli-
cation of (7).
The ellipsoid divides the space R3 into two regions. The
interior of the ellipsoid is:
 
I = x ∈ R3 | (x − a)T B−1 (x − a) < 1 (9)
On the other hand, the exterior of the ellipsoid is:
 
O = x ∈ R3 | (x − a)T B−1 (x − a) > 1 (10)
Next these concepts are applied to robust fitting.
3.2 Robust Fitting
Let us consider a finite set of M input data points in R3 :
 
S = xi ∈ R3 | i ∈ {1, . . . , M} (11)

(x − a)T B−1 (x − a) = 1 (3)
where a ∈ R3 is the center of the ellipsoid (a 3-dimensional
vector), and B ∈ R3×3 is a 3 × 3 positive definite matrix. The
eigenvectors of B are the major axes of the ellipsoid, while
their associated eigenvalues are the squares of the lengths of
these major axes. It must be noted that the ellipsoid surface
Please note that no normal vectors are provided as inputs,
only the data points. The task is to fit an ellipsoid to the input
points in S.
Here we propose to use the least absolute error rather than
the least squared error, so as to obtain a robust performance
criterion [3]. Our proposed error measure is:


1 
M
is the unit level set f (x) = 1 of this function: E= I (xi ∈ O) T −1
(xi − a) B (xi − a) − 1
M
 i=1
f (x) = (x − a)T B−1 (x − a) 

1 
(4) M
T −1
+ I (xi ∈ I ) 1 − (xi − a) B (xi − a)
M
The ellipsoid can be obtained by affine transformation i=1
from the unit sphere with center at the coordinate origin, +λ a (12)
whose equation is:
where λ is a positive weighting parameter for the last term
y y=1
T
(5) which ensures that the solution does not go to infinity, i.e.,
a → ∞, and I is the indicator function:
This is achieved by application of the following affine 
transform: 1 iff condition is true
I (condition) = (13)
0 iff condition is false

x = Ly + a (6) It must be noted that a ellipsoid with a far enough center

a can fit any finite set of points; hence, the need for the last
where L is a 3 × 3 lower triangular matrix obtained by term of (12). Also, it must be highlighted that the definitions
Cholesky decomposition of B; please note that B admits such of the interior set (9) and the exterior set (10) are referred to
decomposition because it is positive definite: the current estimation of the ellipsoid parameters B and a, so
both sets are updated as the estimation changes. This implies
B = LLT (7) that all the terms of (12) are nonnegative, so that E ≥ 0. In
turn this means that (12) is equivalent to:
In order to see this, from (6) we have: M  
1   
E=  (xi − a) B (xi − a) − 1 + λ a
T −1
(14)
y = L−1 (x − a) (8) M
i=1

123
J Math Imaging Vis (2017) 58:189–210 193

which expresses E as the mean of the absolute errors plus In this work we propose to estimate the error trend in
the weighted normalization term. a robust way. To this end, at time instant t the median of
Gradient descent is used to minimize E. Consequently, the errors at time instants t, t − 1, . . . ., t − T is compared
the gradient with respect to B is computed: to the median of the errors at time instants t − T − 1, t −
T − 2, . . . ., t − 2T , and a binary variable ξ (t) ∈ {−1, 1} is
1  1 − 1
M
∂E obtained:
(xi − a)T B−1 (xi − a)
2
=− I (xi ∈ O)
∂B M 2
i=1
Ê new (t) = median (E (t) , E (t − 1) , . . . ., E (t − T ))
×B−T (xi − a) (xi − a)T B−T
(20)
1  1 − 1
M
(xi − a)T B−1 (xi − a)
2
+ I (xi ∈ I ) Ê old (t) = median (E (t − T − 1) , E (t − T − 2) ,
M 2
i=1 . . . ., E (t − 2T )) (21)
×B −T
(xi − a) (xi − a)T B−T (15) 
−1 iff Ê new (t) < Ê old (t)
ξ (t) = (22)
where we note: 1 iff Ê new (t) ≥ Ê old (t)
 T
B−T = B−1 (16) where ξ (t) = −1 indicates that the error exhibits a decreas-
ing trend.
Next the gradient with respect to a is obtained as follows: Then the step size is updated according to the trend ξ (t):
⎧
1 
M
∂E a ⎨ρup η (t) iff ξ (t) = −1 and η (t) < ηmax
=λ − I (xi ∈ O)
∂a a M η (t + 1) =
⎩ρ
down η (t) iff ξ (t) = 1 and η (t) > ηmin
i=1
 − 1
−1 2 −1
× (xi − a) B T
(xi − a) ×B (xi − a) (23)
 − 1 0 < ρdown < 1 (24)
1 
M
I (xi ∈ I ) (xi − a)T B−1 (xi − a)
2
+ ρup > 1 (25)
M
i=1 ηmin < ηmax (26)
−1
×B (xi − a) (17)
where ρup and ρdown control the size of the update of η, and
Consequently, the gradient method yields the following
ηmin and ηmax set limits for such update. It has been found in
update equations:
practice that by taking T = 5, ρup = 1.1 and ρdown = 0.9 the
∂E step size adaptation procedure is stabilized, so these values
B (t + 1) = B (t) − η (t) (t) (18) have been used in all the reported experiments. In order to
∂B
∂E save computation time and improve stability, the above step
a (t + 1) = a (t) − η (t) (t) (19) size update procedure is only carried out when t is a multiple
∂a
of 2T .
where t is the time instant and η (t) is the step size.
3.4 Initialization
3.3 Robust Convergence Control
Here the initialization procedure to obtain the starting values
In order to achieve a faster convergence, an adaptive step
B (0) and a (0) for t = 0 is described. It is necessary to
size η (t) is used which depends on the time instant t. The
produce an initial value of B (0) which is a positive definite
step size should increase when the error decreases in order to
matrix, as indicated in Sect. 3.1. To this end, we take a (0)
advance faster to the local minimum, while it should decrease
as the mean of the training data, and B (0) as the covariance
when the error increases in order to avoid further degradation
matrix of the data:
of the current solution due to wrong updates [46,49,71]. A
potential source of instability in the fitting algorithm is that
1 
M
the error could exhibit a large variability within a given region a (0) = xi (27)
of the search space. This could lead to wrong updates of the M
i=1
step size, which in turn would increase the variability of the
1 
M
error. Therefore, the stability of the method can be enhanced B (0) = (xi − a (0)) (xi − a (0))T (28)
by computing a robust estimation of the current error. M
i=1

123
194 J Math Imaging Vis (2017) 58:189–210

Fig. 1 Four examples of the L1 linear programming fit using the general equation (1) versus our proposal. N = 100 samples were used for the
simulations, percentage of added outlier data P = 10% and Gaussian noise level for inliers σ = 0.01

Equations (27) and (28) ensure that B (0) is positive defi- Table 1 Quadratic surface types obtained by the L1 linear program-
nite except in case that a degenerate training set is provided ming fit using the general equation (1) over 100 simulations
which lies in a plane. In case that a more robust initialization Type of quadratic surface %
is necessary, robust covariance matrix estimation methods are
Hyperboloid of one sheet 66
available [7,29,77]. They have not been used here because
Hyperboloid of two sheets 10
the above procedure yields good results with a reduced com-
putational complexity. Ellipsoid (real) 24

123
J Math Imaging Vis (2017) 58:189–210 195

Coefficient errors 7. If an error goal or an iteration limit have been reached,

2
then go to step 8. Otherwise, go to step 2.
1 8. Output the best found solution, i.e., the values of a and
B associated to the minimum recorded value of E, and
0
log 10 (e coeff)
then halt.
-1
3.6 Discussion
-2

-3 In this subsection some important aspects of our proposal

are discussed. In particular, the differences of our approach
-4 with respect to previous methods are considered, and some
L1 linear fit Proposed fit
directions for future research are outlined.
Fig. 2 Coefficient errors of the L1 linear programming fit and the pro- First of all, it must be studied why standard methods for
posed fit. Boxplots of the 2-norm errors between the true and the fitted L1-norm minimization cannot be employed in our approach.
coefficients of Eq. (1) over 100 runs are displayed (in a logarithmic
One of the best known approaches to replace least squares
scale). In the boxplots, the medians are plotted as horizontal gray lines,
while the means are plotted as gray circles minimization by a Lp-norm is the iteratively reweighted least
squares (IRLS) method [13,74]. It employs specially selected
weights for the input samples so that Lp-norm minimization
3.5 Summary can be obtained by solving a weighted least squares problem.
It has been successfully applied to various problems in image
The proposed algorithm can be summarized as processing and computer vision [11,17,51]. However, IRLS
follows: is only suitable for problems with a linear objective function.
In our case, the proposed objective function (14) is nonlinear,
which precludes the use of IRLS.
1. Set the initial values of a and B according to (27) and The nonlinearity of the objective function also excludes
(28), respectively. linear programming as an alternative. It should be noted that
2. Record the value of the error E for the current solution. the general equation of 3D quadratic surfaces (1) is linear in
3. Classify the training data into interior or exterior points the 10 parameters a, b, c, d, f , g, h, p, q, r . Therefore, it
with (9) and (10). can be used to carry out a L1 norm optimization by linear
4. Find ∂∂BE and ∂∂aE by equations (15) and (17), respectively. programming. However, the best fit to the data might not be
5. Update the current values of B and a according to (18) an ellipsoid because (1) describes other kinds of surfaces. In
and (19), respectively. order to illustrate this point, Fig. 1 depicts an example of the
6. If the time instant t is a multiple of 2T , then update the results of the L1 linear programming fit compared with the
step size according to (23). true ellipsoid given by synthetic dataset and also the output

Fig. 3 Results of our algorithm 2.5

with perturbed initial
estimations of the ellipsoid 0.6
2
parameters, generated by the
0.5
addition of Gaussian noise with
standard deviation ρ: a error 0.4 1.5
E

measure E versus time step t, b

learning rate η versus time step 0.3
1
t, c legend
0.2
0.5
0.1

0 0
0 1 2 3 4 5 0 0.5 1 1.5 2
log 10 (t) t 10 4

(a) (b)

= 0.01 = 0.05 = 0.1 = 0.25 = 0.5 =1

(c)

123
196 J Math Imaging Vis (2017) 58:189–210

Center position Distance matrix Volume

4 5
2
3 4

1
2 3
log 10 (e a )

log 10 (e B )

log 10 (e V )
0 2
1
1
-1 0
0
-2 -1
-1
50 100 150 200 250 50 100 150 200 250 50 100 150 200 250
N N N
(a) (b) (c)
First principal axis length Angular orientation Running time
5 60

4 50 0

log 10 (CPU time)

3 40
log 10 (e )

-1
2 30
e

1 20 -2

0 10
-3
-1 0
50 100 150 200 250 50 100 150 200 250 50 100 150 200 250
N N N
(d) (e) (f)
ours direct koopmans distance leastsquares simple

Fig. 4 Comparison of the performance metrics of all methods depend- length error, e median of the orientation errors for the three ellipsoid
ing on the number of data points. Fixed values σ = 0.01 and P = 10% axes (degrees), f running time (seconds). In the boxplots, the medians
were used as N was varied. Boxplots of 100 runs are displayed: a cen- are plotted as horizontal gray lines, while the means are plotted as gray
ter error, b distance matrix error, c volume error, d first principal axis circles

of our proposed algorithm. As seen, L1 norm optimization
yields non-ellipsoidal surfaces in many cases, as equation
(1) represents 17 types of quadratic surfaces. In Table 1 the
surface types obtained in 100 runs with different random
samples are reported. We can observe that 76% of the runs
yield an hyperboloid of one or two sheets, while an ellipsoid
is obtained in only 24% of the cases. Moreover, when an
ellipsoid is obtained the fit is far worse than ours, as seen in
Fig. 1(b). The 2-norm of errors ecoeff in the coefficients of
the general equation between the true ellipsoid and the fitted
surface were calculated for both the L1 linear programming
fit and the proposed method:
α = (a, b, c, d, f, g, h, p, q, r )
ecoeff = ||αtrue − αfit ||
Figure 2 displays the boxplots of those errors for the 100
runs (in a logarithmic scale). There is a large difference in
the coefficient errors, which means that the proposed method
outperforms the L1 linear programming fit.
Quadratic programming and other more general
approaches such as semidefinite programming [72] cannot be
applied either, because our objective function (14) involves
the square root of a nonlinear function of the unknown param-
eters of the ellipsoid B and a.
Next some possible extensions of our approach are out-
lined. It must be highlighted that the framework presented
here could be extended to other error norms by changing the
error measure E in equation (12), given a method to compute
the gradients ∂∂BE and ∂∂aE . It might be applied to fit hyperel-
lipsoids in dimensions higher than 3, since the derivations in
Sects. 3.1 and 3.2 are still valid in Rn for n ≥ 3. In princi-
ple, it could be extended to other quadratic forms or higher
(29)
degree polynomials in implicit form F (x) = 0 by defining
(30) two regions F (x) > 0 and F (x) < 0. In this case, the gradi-
1 1
ents of (F (x)) d and (−F (x)) d with respect to the surface
parameters would be required, respectively, where d is the
degree of the polynomial F.

123
J Math Imaging Vis (2017) 58:189–210 197

Center position Distance matrix Volume

3 5
4
2 4
3
1 3
log 10 (e a )

log 10 (e B )

log 10 (e V )
2
0 2
1 1
-1
0 0
-2
-1 -1
-3
-2 -2
0.001 0.005 0.01 0.05 0.1 0.001 0.005 0.01 0.05 0.1 0.001 0.005 0.01 0.05 0.1

(a) (b) (c)

First principal axis length Angular orientation Running time
5 60

4 50 0

log 10 (CPU time)

3
40
log 10 (e )

2 -1
30
e

1
20 -2
0

-1 10
-3
-2 0
0.001 0.005 0.01 0.05 0.1 0.001 0.005 0.01 0.05 0.1 0.001 0.005 0.01 0.05 0.1

(d) (e) (f)

ours direct koopmans distance leastsquares simple

Fig. 5 Comparison of the performance metrics of all methods depend- principal axis length error, e median of the orientation errors for the
ing on the Gaussian noise level for the inliers. Fixed values N = 100 three ellipsoid axes (degrees), f running time (seconds). In the box-
and P = 10% were used as σ was varied. Boxplots of 100 runs are plots, the medians are plotted as horizontal gray lines, while the means
displayed: a center error, b distance matrix error, c volume error, d first are plotted as gray circles

4 Experimental Results 4.1 Synthetic Data

In this section computational experiments for different data
sets are reported.1 The following competing methods have
been chosen for comparison:
– Direct Method (DM) [40],
– Least Square (LS), [32],
– Koopmans method, [70],
– Distance, [32] and
– Simple [40].
Three kinds of data have been employed: synthetic data
(Sect. 4.1), data from a 3D scanner (Sect. 4.2), and finally
stereo image pairs from a calibrated stereo camera (Sect. 4.3).
1 The source code and some demos of our approach are published in
http://www.lcc.uma.es/%7Eezeqlr/ellipsoid/ellipsoid.html.
As a first experiment, some synthetic data with added outlier
points are used to compare to previous methods. In order to
study how our algorithm depends on the initial estimation of
the ellipsoid parameters, an experiment has been carried out
which perturbs the initial estimations of the center a and the
distance matrix B through the addition of Gaussian noise to
the training samples used to compute a (0) and B (0), with
standard deviations ρ = 0.01, 0.05, 0.1, 0.25, 0.5, 1. Please
note that the training samples for the subsequent steps of the
algorithm are not corrupted by this Gaussian noise.
The error measure E (equation 14) and the learning rate
η (step size) along the execution steps of the algorithm have
been recorded. The results are shown in Fig. 3. These results
demonstrate that regardless of the initial estimation of the
ellipsoid, the final measured error is almost the same, even
in the case of large initial perturbations. It must be noted that

123
198 J Math Imaging Vis (2017) 58:189–210

Center position Distance matrix Volume

5
8 8
4

3 6 6
log 10 (e a )

log 10 (e B )

log 10 (e V )
2
4 4
1

0 2 2

-1
0 0
-2

0.05 0.1 0.15 0.2 0.25 0.05 0.1 0.15 0.2 0.25 0.05 0.1 0.15 0.2 0.25
P P P
(a) (b) (c)
First principal axis length Angular orientation Running time
60
8
50 0

log 10 (CPU time)

6
40
log 10 (e )

-1
4 30
e

2 20 -2

0 10
-3
0
0.05 0.1 0.15 0.2 0.25 0.05 0.1 0.15 0.2 0.25 0.05 0.1 0.15 0.2 0.25
P P P
(d) (e) (f)
ours direct koopmans distance leastsquares simple

Fig. 6 Comparison of the performance metrics of all methods depend- length error, e median of the orientation errors for the three ellipsoid
ing on the percentage of outliers. Fixed values N = 100 and σ = 0.01 axes (degrees), f running time (seconds). In the boxplots, the medians
were used as P was varied. Boxplots of 100 runs are displayed: a cen- are plotted as horizontal gray lines, while the means are plotted as gray
ter error, b distance matrix error, c volume error, d first principal axis circles

initial values of E are lower for higher values of ρ. This is

a = a − atrue  (31)
because higher noise ρ produces a covariance matrix of the
B = B − Btrue  (32)
data B (0) with a higher value of det (B (0)) due to the higher π 3/2 
dispersion of the data. This implies that (x − a)T B−1 (x − a) V = 3  det (B) (33)
is smaller during the first steps of the algorithm and, hence,
E is smaller. The learning rate η is not significantly affected
V = |V − Vtrue |
by the perturbation in the initial estimation of the ellipsoid,
λ = |λ − λtrue |
which indicates that the algorithm takes steps of similar size.
θ = mediani∈{1,2,3} v
Hence, the result of the iterative procedure is not degraded
under perturbations of the initial ellipsoid parameter estima-
tions.
The second experiment with synthetic data comprises a set
of repeated simulation tests to compare the performance of
the competing approaches. In order to carry out quantitative
comparisons with the competing methods, the center error
(31), the distance error (32), the volume error (34), the major
axis length error (35) and the angular orientation error of the
principal axes (36) with respect to the true (ground truth)
ellipsoid have been computed:
123
 2 +1
(34)
(35)
i vtrue,i (36)
where the norm · in (31) stands for the Euclidean norm
of a vector, the norm · in (32) stands for the 2-norm of a
matrix, λ indicates the principal eigenvalue of B in (35) and
vi is the eigenvector associated to the i-th largest eigenvalue
in (36).
Three kinds of tests with synthetic data have been consid-
ered, by varying the following parameters:
1. The number of inlier training data points randomly drawn
from the ellipsoid surface: N = 50, 100, 150, 200, 250.
J Math Imaging Vis (2017) 58:189–210 199

ours direct koopmans

Training samples Training samples Training samples
True ellipsoid True ellipsoid True ellipsoid
Our method direct koopmans
10 10 15

10
5 5

5
0 0
0

-5 -5
-5

-10 -10 -10

10 10 10
10 10 20
5 5 5
10
0 0 0 0 0
0
-5 -10 -5 -10 -5 -10

(a) (b) (c)

distance leastsquares simple

Training samples Training samples Training samples
True ellipsoid True ellipsoid True ellipsoid
distance leastsquares simple
15 10 15

10 10
5

5 5
0
0 0

-5
-5 -5

-10 -10 -10

10 10 10
5 20 10 20
5 5
0 10 10
0 0 0
-5 0 0
-10 -10 -5 -10 -5 -10

(d) (e) (f)

Fig. 7 Solutions for synthetic data with outliers. a Our method. b Direct method. c Koopmans method. d Distance method. e Least Square method.
f Simple method

Table 2 Quantitative errors for example synthetic dataset

a
B
V
λ
θ1
θ2
θ3 Runtime (s)

Ours 0.048975 0.378138 1.683371 0.011273 3.056489 3.058897 0.224226 3.109880

Direct 2.160242 15.301255 136.342838 0.136268 52.008192 56.020176 11.029736 0.103589
Koopmans 3.276190 178.247165 243.922386 6.891759 51.217157 59.683413 15.304854 0.121484
Distance 4.608043 107.567183 434.093471 4.225820 42.076599 63.587081 19.304151 0.448471
Least Square 2.206723 57.779994 289.667286 2.329665 54.819331 58.803372 14.229513 0.007271
Simple 4.208742 230.711858 311.385268 8.521828 53.669883 58.012052 13.494440 0.001149

2. Additive Gaussian noise for the inliers, with σ = 0.001,
0.005, 0.01, 0.05, 0.1.
3. Percentage of added outliers in the training data: P = 5,
10, 15, 20, 25%.
Please note that the overall number of training data points
M is the sum of the inlier and outlier points, i.e., M = N +
P has been randomly generated and provided as input to all the
100 N . The inlier generation process is as follows. First of all,
N random samples on the unit sphere are generated. Then
Gaussian noise is added to those samples. Finally, the unit
sphere is linearly transformed to yield an ellipsoid. Therefore,
σ = 0.1 means that the standard deviation of the noise is 10%
of the ellipsoid axis length in each principal direction. For an
object of 1 meter this corresponds to a standard deviation of
the noise of 10 centimeters.
For each kind of test and value of the varying parameter,
100 simulation runs have been carried out to yield significant
comparisons. For each simulation run, a new training dataset
competing algorithms. The following fixed values were used
as one of the parameters was varied: N = 100, σ = 0.01,
P = 10%.

123
200 J Math Imaging Vis (2017) 58:189–210

Fig. 8 Real scanned object

with few outliers: a 3D scanned
points, b evolution of the center
as estimated by our method
60

50

40

30

20

10

0
40
20 40
0 20
0
-20
-20
-40 -40

(a) (b)
ours direct koopmans

Training samples Training samples Training samples

Our method direct koopmans

60 60 60

50 50 50

40 40 40

30 30 30

20 20 20

10 10 10

0 0 0

20 20 20
10 20 10 20 10 20
0 0 0
0 0 0
-10 -10 -10
-20 -20 -20 -20 -20 -20

(a) (b) (c)

distance leastsquares simple

Training samples Training samples Training samples

distance leastsquares simple

60 60 60

50 50 50

40 40 40

30 30 30

20 20 20

10 10 10

0 0 0

20 20 20
10 20 10 20 10 20
0 0 0
0 0 0
-10 -10 -10
-20 -20 -20 -20 -20 -20

(d) (e) (f)

Fig. 9 Solutions for scanned real object with few outliers. a Our method. b Direct method. c Koopmans method. d Distance method. e Least
Square method. f Simple method

Three rounds of simulations have been carried out, fixing
two of the parameters and varying the other. For each value
of each varying parameter and each competing method, we
have obtained the mean, median and quartiles over the set of
100 runs, and we have plotted their respective boxplots. The
performance of the competing methods has been evaluated
by comparison with the ground truth ellipsoid in terms of the
metrics explained before. Also, we obtained the runtime of all
methods on a desktop personal computer with a 64-bit quad-
core 3GHz CPU and standard hardware. All the methods are
implemented as MATLAB scripts, with no use of the GPU.
Figure 4 shows the comparisons among methods by vary-
ing the number of samples N . As shown in the figure, where
boxplots of the runs are displayed, the center, distance matrix

123
J Math Imaging Vis (2017) 58:189–210
Table 3 Volume errors for the real scanned object with few outlier
points
V center, distance matrix and volume errors. Our method is less
Ours 67,818.428852
Direct 67,978.753202
Koopmans 67,905.470799
Distance 67,825.003952
Least Square 68,091.613165
Simple 67,971.679099
and volume error for our proposal are the lowest (note that
these figures have a logarithmic scale). It is followed by
Direct, Least Squares and Simple algorithms, whose ten-
dency is the same as ours, i.e., decreasing error as more
data samples are employed. However, Koopmans and Dis-
tance methods have a different behavior, namely obtaining
worse results as N is increased. Values near to zero of the
principal axes (length and orientation) confirm that the fit-
ted ellipsoid from our method is the best approximation to
the true ellipsoid. Overall it can be observed that our errors
are almost always lower than those of the other methods
and decrease when more data points are employed, which
indicates convergence to the real ellipsoid. Also, outliers are
almost absent in the performance results, while the mean and
median, which are displayed as an horizontal line and a cir-
cle, respectively, have almost the same value, indicating a
high degree of robustness. This does not happen in the Dis-
tance, Koopmans or Simple algorithms. The best results in
the execution time are achieved by Direct and Least squares
methods, but they are less stable. Our algorithm is not signif-
icantly affected in its complexity by random data variations,
i.e., its runtime is stable across simulation runs for a given
training set size.
Figure 5 shows the comparisons varying the Gaussian
noise level σ for the inlier data points. At low noise lev-
els, although the data are more dispersed, our approach has a
Fig. 10 First real scanned
object with many outlier points:
a 3D scanned points, b
evolution of the center as
estimated by our method
(a)
201
better behavior than the other methods, because it is the only
one which has negative values in the logarithmic scales of the
efficient in the orientation, but it is still the best one, attaining
mean errors no larger than 12o with a very small interquartile
range. In Fig. 5 it can be seen that competitors have higher
errors even at small values of σ , but when the noise raises,
the differences among the methods are smaller. The order of
the methods remains the same as when the number of data
samples is modified. Furthermore, the running time is not
affected by the noise.
Figure 6 depicts the results of the comparisons vary-
ing the percentage of outliers P added to the data set. In
general terms, when there are more outliers, worse results
are obtained. However, the precision of our method is the
best one. If there are more than 15% of added outliers, our
results in the distance matrix and volume errors are close
to the Direct, Least Squares and Simple algorithms. Nev-
ertheless, the results of the center errors and the axis length
and angular orientation, and especially their lower dispersion,
allow concluding that our method is the best performing one.
Regarding the median in Fig. 6 it can be seen that competitors
have an error of more than 102 in the volume and also the
distance matrix with respect to the true ellipsoid. The center
and the first principal axis length are also very different. The
huge interquartile range in the boxplots of the angular orien-
tation reveals that the rest of the methods are very imprecise
when compared with our approach. Again our execution time
is slower, around 3 seconds, which is the same as in the other
tests.
In order to provide a qualitative assessment of the results
with synthetic data, Fig. 7 depicts an example of the results of
the competing methods for a given synthetic dataset. It cor-
responds to N = 100 samples, percentage of added outlier
data P = 10%, and Gaussian noise level for inliers σ = 0.01.
The associated quantitative performance metrics are shown
in Table 2. As seen, our method clearly outperforms all its
competitors.
80
60
40
20
0
50
40
20
0 0
-20
-40
-50 -60
(b)
123
202 J Math Imaging Vis (2017) 58:189–210

ours direct koopmans

Training samples Training samples Training samples

Our method direct koopmans

60 60 60

40 40 40

20 20 20

0 0 0

-20 -20 -20

40 40 40
20 20 20 20 20 20
0 0 0 0 0 0
-20 -20 -20 -20 -20 -20
-40 -40 -40
-40 -40 -40

(a) (b) (c)

distance leastsquares simple

Training samples Training samples Training samples

distance leastsquares simple

60 60 60

40 40 40

20 20 20

0 0 0

-20 -20 -20

40 40 40
20 20 20 20 20 20
0 0 0 0 0 0
-20 -20 -20 -20 -20 -20
-40 -40 -40
-40 -40 -40

(d) (e) (f)

Fig. 11 Solutions for the first real scanned object with many outliers. a Our method. b Direct method. c Koopmans method. d Distance method.
e Least Square method. f Simple method

Table 4 Volume errors for the first real scanned object with many out-
lier points
V true center and the true distance matrix are unknown.
Ours
Direct
Koopmans
Distance
Least Square
Simple
4.2 Scanner Data
The second round of experiments considers real objects
analyzed by a 3D scanner. The real data used for these exper-
iments have been acquired by using the Roland LPX 1200
3D laser Picza™ scanner with the following characteristics:
noncontact laser sensor, spot-beam triangulation scanning
method, repeat accuracy of ±0.05 mm and laser wavelength
from 645 to 660 nm. It must be highlighted that the quan-
titative comparisons for real objects can only be done with
respect to the volume error (34) because the scanner only
provides the true volume of the scanned object, while the
Three pieces of meal with near ellipsoidal shapes have
65,659.271694
been chosen for these experiments, one with few outliers
115,980.693801
and two with many outliers and significant occlusion. As it is
1,316,316.681929 shown in Fig. 8 we chose an egg as a 3D object with few out-
2,065,415.137695 lier points. Results are quite similar without any significant
166,113.326337 differences among the applied methods, as seen in Fig. 9.
123,889.885783 In Table 3 the volume error among the studied methods is
shown. Again, no significant differences are noticed.
Another experiment to compare these methods is based on
scanning a 3D object with many outlier points and a signifi-
cant occluded section. In Fig. 10, a real object is shown with
a dense cloud of outlier points. Figure 11 shows the fitting
result of the different methods. The robustness of our method
is displayed in Fig. 11a. Table 4 reports the comparison of the
volume errors using these real data with many outliers. Our
approach is shown to consistently outperform its competitors
both in qualitative and quantitative terms.
As the last real dataset a scanned piece of fruit with many
outlier points and occlusion has been tested (see Fig. 12).

123
J Math Imaging Vis (2017) 58:189–210
Fig. 12 Second real scanned
object with many outlier points:
a 3D scanned points, b
evolution of the center as
estimated by our method
(a)
Table 5 Volume errors for the second real scanned object with many
outlier points
V [42], with the maximum threshold on the forward–backward
Ours 90,409.475127
Direct 144,565.170593
Koopmans 161,886.876864
Distance 171,374.993323
Least Square 173,355.609190
Simple 154,104.187763
In Table 5 the ellipsoid fitting volume error comparison is
shown, which reveals that our method’s estimated volume
is much closer to the real one than those of the competing
approaches. Finally, Fig. 13 shows the qualitative results.
Once again, our method attains a close approximation to the
real ellipsoid.
4.3 Stereo Camera Data
The third round of experiments deals with stereo image pairs
acquired by a calibrated stereo camera, so that the fitted ellip-
soids can be superimposed on the acquired RGB images.
These experiments are challenging since about one half of all
objects are not visible, i.e., approximately 50% of the object
surfaces are missing from the input data sets. The camera
employed for these experiments is a JVC GS TD1™. First
of all, the camera was calibrated by Zhang’s method [78]
to obtain the intrinsic camera parameters. Backprojection
errors lower than 0.5 pixels were attained. After that pairs
of images were taken of approximately ellipsoidal objects
on a neutral background. The images had 1920 × 1024
pixel resolution and 8 bit precision for each RGB chan-
nel. For each image pair, feature points were searched in
the left and right rectified images by the minimum eigen-
value algorithm [60], with minimum quality parameter 0.001
203
80
60
40
20
0
40
20 20
0
0 -20
-20 -40
-60
-40 -80
(b)
and Gaussian filter size 5 × 5. Then stereo point pairs were
identified by the Kanade-Lucas-Tomasi (KLT) algorithm
error set to 1, and 5 pyramid levels. The stereo pairs were
located on the intrinsic 3D coordinate system of the cam-
era by triangulation, and finally the resulting 3D data points
were provided to the competing ellipsoid fitting algorithms.
After the estimated ellipsoids were computed, they were
backprojected to both the left and right images for visual
assessment. For a better comparison, we have only con-
sidered the two methods that have the best performance in
addition to the proposed one, namely the Direct and Simple
methods.
Figure 14 shows a tridimensional reconstruction of a bas-
ketball and the adjustment of each method. The Direct and
Simple methods obtain an ellipsoid with a similar shape to
the basketball, but underestimate the dimensions. This can be
better appreciated in Fig. 15, where the ellipsoids are super-
imposed on the right stereo image. Our fitted ellipsoid covers
most of the ball and is centered, which means that our method
is less affected by the outliers on the right side of Fig. 14,
while the other methods yield displaced and more stretched
ellipsoids.
Figure 16 shows a tridimensional reconstruction of a
loquat and the adjustment of each method. The three meth-
ods obtain ellipsoids of almost the same size, but our method
yields a better estimation of the orientation. Figure 17 depict
the superimposition on the right stereo image. There it can be
seen that our method produces an ellipsoid with the same ori-
entation as the loquat, while the others are slightly displaced
to the right due to the presence of outliers.
Figure 18 shows a tridimensional reconstruction of a
balloon and the adjustment of each method. Although
Koopmans, Least Squares and Distance algorithms are not
displayed here, they obtain larger ellipsoids. This might be
caused by the irregular shape of the balloon, which has two
protuberances on the right side. Direct and Simple methods
123
204 J Math Imaging Vis (2017) 58:189–210

ours direct koopmans

Training samples Training samples Training samples

Our method direct koopmans
80 80 80

60 60 60

40 40 40

20 20 20

0 0 0
40 40 40
20 50 20 50 20 50
0 0 0 0 0 0
-20 -50 -20 -50 -20 -50
-40 -100 -40 -100 -40 -100

(a) (b) (c)

distance leastsquares simple

Training samples Training samples Training samples

distance leastsquares simple
80 80 80

60 60 60

40 40 40

20 20 20

0 0 0
40 40 40
20 50 20 50 20 50
0 0 0 0 0 0
-20 -50 -20 -50 -20 -50
-40 -100 -40 -100 -40 -100

(d) (e) (f)

Fig. 13 Solutions for the second real scanned object with many outliers. a Our method. b Direct method. c Koopmans method. d Distance method.
e Least Square method. f Simple method

4 4 4
10 10 10
2 2 2
z-axis

z-axis

z-axis

1.5 1.5 1.5

1 1 1

3000 3000 3000

2000 2000 2000

y-axis

y-axis

y-axis

1000 1000 1000

0 0 0

-1000 -1000 -1000

12 12 12
-2000 -2000 -2000
-2000 -1000 -2000 -1000 -2000 -1000
0 0 0
1000 1000 1000
2000 2000 2000
x-axis x-axis x-axis

(a) (b) (c)

Fig. 14 Solutions for our method and Direct and Simple methods using a 3D stereo reconstruction of a basketball with some outliers. The position
of the left lens is marked as ‘1’, and the right lens is marked as ‘2’. a Our method. b Direct method. c Simple method

123
J Math Imaging Vis (2017) 58:189–210 205

200 200 200

400 400 400

600 600 600

800 800 800

1000 1000 1000

500 1000 1500 500 1000 1500 500 1000 1500

(a) (b) (c)

Fig. 15 Projections on the right image of basketball of the solutions for our method and Direct and Simple methods. a Our method. b Direct
method. c Simple method

4 4 4
10 10 10

2 2 2

1.5 1.5 1.5

z-axis

z-axis

z-axis
1 1 1

0.5 0.5 0.5

4000 4000 4000
0 0 0
2000 12 2000 2000 12 2000 2000 12 2000
0 0 0
-2000 0 y-axis -2000 0 y-axis -2000 0 y-axis
x-axis -4000 x-axis -4000 x-axis -4000

(a) (b) (c)

Fig. 16 Solutions for our method and Direct and Simple methods using a 3D stereo reconstruction of a loquat with some outliers. The position of
the left lens is marked as ‘1’, and the right lens is marked as ‘2’. a Our method. b Direct method. c Simple method

200 200 200

400 400 400

600 600 600

800 800 800

1000 1000 1000

500 1000 1500 500 1000 1500 500 1000 1500

(a) (b) (c)

Fig. 17 Projections on the right image of a loquat of the solutions for our method and Direct and Simple methods. a Our method. b Direct method.
c Simple method

123
206 J Math Imaging Vis (2017) 58:189–210

12000 12000 12000

10000 10000 10000

z-axis

z-axis

z-axis
8000 8000 8000

6000 6000 6000

-1000 -1000 -1000

-500 -500 -500

0 0 0
y-axis

y-axis

y-axis
500 500 500

1000 1000 1000

2 1 3000 2 1 3000 2 1 3000
1500 1500 1500
2000 2000 2000
1000 1000 1000
2000 0 2000 0 2000 0
-1000 x-axis -1000 x-axis -1000 x-axis

(a) (b) (c)

Fig. 18 Solutions for our method and Direct and Simple methods using a 3D reconstruction of a balloon with some outliers. The position of the
left lens is marked as ‘1’, and the right lens is marked as ‘2’. a Our method. b Direct method. c Simple method

200 200 200

400 400 400

600 600 600

800 800 800

1000 1000 1000

500 1000 1500 500 1000 1500 500 1000 1500

(a) (b) (c)

Fig. 19 Projections on the right image of a balloon of the solutions for our method and Direct and Simple methods. a Our method. b Direct method.
c Simple method

obtain better results, but they are affected by the presence Algebraic criteria lead to efficient algorithms to the fit-
of outliers (lower part of the image), which causes a dis- ting problem but their accuracy can be unacceptable in the
placement of the estimated ellipsoid. However, this does not presence of outliers. On the other hand, traditional geomet-
happen with our algorithm. Figure 19 shows the superimposi- ric criteria that are used in fitting an ellipsoid are based
tion on the right stereo image. It can be seen that our method on a differentiable objective function. They give fast iter-
produces the best fit. Direct and Simple yield a displaced ative algorithms based on gradients or second derivatives,
center, which it does not occur for our proposal. using nonlinear programming methods. However, these cri-
teria could be very sensitive to outliers or data noise.
Experimental results show that our algorithm is robust
5 Conclusions since the solutions have a low sensitivity to outliers and occlu-
sion. The robustness of the proposed algorithm is due to the
In this paper, we have developed an algorithm to fit an usage of absolute rather than squared errors, since the sum
ellipsoid to a set of points by minimizing the mean abso- of absolute errors is more robust than the sum of squared
lute deviation from the level set which defines the ellipsoid errors. Furthermore, the adaptive step control procedure uses
surface. This objective function presents more robust fit the median of the last observed absolute errors; as it is well
results than other compared methods. The proposed algo- known the median is a robust estimator. Finally, the solution
rithm requires an iterative search using a positive definite ellipsoid attains a balance among the outside points and the
distance matrix to fit the feature. Nevertheless, independently inside points. These algorithm features are responsible for its
of the chosen initial parameters the proposed algorithm con- robustness against outliers and occlusion. As seen in the sim-
verges to a finite solution. ulation tests with synthetic data, our approach consistently

123
J Math Imaging Vis (2017) 58:189–210
outperforms its competitors in all performance metrics, with
the exception of the runtime, where our algorithm is slower.
Also, a substantially smaller dispersion in the performance
results are found for our method, which indicates that our
approach is more reliable. The experiments with real data
from a 3D scanner and a stereo camera confirm the results
obtained for synthetic data, even for objects with significant
occlusion.
Some of the main computational advantages of proposed
algorithm are that the objective function usually decreases at
each iteration because it uses an adaptive step, the number of
iterations is usually reduced, and the operations performed
by the algorithm are mainly based on the sum of the abso-
lute errors coming from the distances from each point to the
center.
Therefore, it can be concluded that a competitive approach
for ellipsoid fitting has been proposed, which can be used for
datasets with significant amounts of outliers and occlusion.
This enhances its applicability to real problems with low
quality input data.
Acknowledgements This work is partially supported by the Ministry
of Economy and Competitiveness of Spain under Grant TIN2014-
53465-R, project name Video surveillance by active search of anoma-
lous events. It is also partially supported by the Autonomous Gov-
ernment of Andalusia (Spain) under projects TIC-6213, project name
Development of Self-Organizing Neural Networks for Information
Technologies; and TIC-657, project name Self-organizing systems and
robust estimators for video surveillance. All of them include funds
from the European Regional Development Fund (ERDF). The authors
thankfully acknowledge the computer resources, technical expertise and
assistance provided by the SCBI (Supercomputing and Bioinformatics)
center of the University of Málaga. They have also been supported by the
Biomedic Research Institute of Málaga (IBIMA). They also gratefully
acknowledge the support of NVIDIA Corporation with the donation
of the Titan X GPU used for this research. Karl Thurnhofer-Hemsi is
funded by a Ph.D. scholarship from the Spanish Ministry of Education,
Culture and Sport under the FPU program.
References
1. Ahn, S.J.: 3. Orthogonal distance fitting of implicit curves and
surfaces. In: Least Squares Orthogonal Distance Fitting of Curves
and Surfaces in Space, Lecture Notes in Computer Science, vol.
3151, pp. 35–53. Springer, Berlin (2004)
2. Blane, M., Lei, Z., Civi, H., Cooper, D.: The 3L algorithm for
fitting implicit polynomial curves and surfaces to data. IEEE Trans.
Pattern Anal. Mach. Intell. 22(3), 298–313 (2000)
3. Branham Jr., R.L.: Alternatives to least squares. Astron. J. 87, 928–
937 (1982)
4. Calafiore, G.: Approximation of n-dimensional data using spherical
and ellipsoidal primitives. IEEE Trans. Syst. Man Cybern. A: Syst.
Hum. 32(2), 269–278 (2002)
5. Cao, X., Shrikhande, N.: Quadric surface fitting for sparse range
data. In: Decision Aiding for Complex Systems, Conference Pro-
ceedings of the 1991 IEEE International Conference on Systems,
Man, and Cybernetics, 1991, vol. 1, pp. 123–128 (1991)
6. Chen, Y., Liu, C.: Quadric surface extraction using genetic algo-
rithms. Comput. Aided Des. 31(2), 101–110 (1999)
207
7. Chenouri, S., Liang, J., Small, C.: Robust dimension reduction.
Wiley Interdiscip. Rev.: Comput. Stat. 7(1), 63–69 (2015)
8. Chernov, N., Lesort, C.: Least squares fitting of circles. J. Math.
Imaging Vis. 23(3), 239–252 (2005)
9. Chung, P., Ng, G.: Comparison between an accelerometer and
a three-dimensional motion analysis system for the detection of
movement. Physiotherapy 98(3), 256–259 (2012). Special Issue
on Advancing Technology including papers from WCPT
10. Collings, S., Kozera, R., Noakes, L.: Robust surface fitting from
two views using restricted correspondence. J. Math. Imaging Vis.
34(2), 200–221 (2009)
11. Comport, A., Marchand, E., Pressigout, M., Chaumette, F.: Real-
time markerless tracking for augmented reality: the virtual visual
servoing framework. IEEE Trans. Vis. Comput. Graph. 12(4), 615–
628 (2006)
12. Cross, G., Zisserman, A.: Quadric reconstruction from dual-space
geometry. In: Sixth International Conference on Computer Vision,
1998, pp. 25–31 (1998)
13. Daubechies, I., Devore, R., Fornasier, M., Güntürk, C.: Iteratively
reweighted least squares minimization for sparse recovery. Com-
mun. Pure Appl. Math. 63(1), 1–38 (2010)
14. Eberly, D.: Geometric Tools, LLC. http://www.geometrictools.
com
15. Fang, J., Liu, Z.: A new inclination error calibration method of
motion table based on accelerometers. IEEE Trans. Instrumen.
Meas. 64(2), 487–493 (2015)
16. Fang, J., Sun, H., Cao, J., Zhang, X., Tao, Y.: A novel calibration
method of magnetic compass based on ellipsoid fitting. IEEE Trans.
Instrum. Meas. 60(6), 2053–2061 (2011)
17. Figueiredo, M., Bioucas-Dias, J., Nowak, R.: Majorization–
minimization algorithms for wavelet-based image restoration.
IEEE Trans. Image Process. 16(12), 2980–2991 (2007)
18. Fitzgibbon, A., Pilu, A., Fisher, R.: Direct least square fitting of
ellipses. IEEE Trans. Pattern Anal. Mach. Intell. 21(5), 476–480
(1999)
19. Flöry, S., Hofer, M.: Surface fitting and registration of point clouds
using approximations of the unsigned distance function. Comput.
Aided Geom. Des. 27(1), 60–77 (2010)
20. Gander, W., Golub, G.H., Strebel, R.: Least-squares fitting of cir-
cles and ellipses. BIT Numer. Math. 34(4), 558–578 (1994)
21. Gang, Y., Yingtang, Z., Hongbo, F., GuoQuan, R., Zhining, L.:
Integrated calibration of magnetic gradient tensor system. J. Magn.
Magn. Mater. 374, 289–297 (2015)
22. Ge, Z., S, B., Chan, H.P., Hadjiiski, L.M., Cascade, P.N., Bogot, N.,
Zhou, C.: Computer-aided detection of lung nodules: false positive
reduction using a 3D gradient field method and 3d ellipsoid fitting.
Robot. Auton. Syst. 8(32), 2443–2454 (2005)
23. Gietzelt, M., Wolf, K.H., Marschollek, M., Haux, R.: Performance
comparison of accelerometer calibration algorithms based on 3D-
ellipsoid fitting methods. Comput. Methods Progr. Biomed. 111(1),
62–71 (2013)
24. Gonzalez-Villanueva, L., Cagnoni, S., Ascari, L.: Design of a wear-
able sensing system for human motion monitoring in physical
rehabilitation. Sensors 13(6), 7735 (2013)
25. Gotardo, P., Bellon, O., Boyer, K., Silva, L.: Range image segmen-
tation into planar and quadric surfaces using an improved robust
estimator and genetic algorithm. IEEE Trans. Syst. Man Cybern.
B: Cybern. 34(6), 2303–2316 (2004)
26. Grammalidis, N., Strintzis, M.: Head detection and tracking by 2D
and 3D ellipsoid fitting. In: Proceedings of the Computer Graphics
International, 2000, pp. 221–226 (2000)
27. Grivon, D., Vezzetti, E., Violante, M.G.: Development of an inno-
vative low-cost MARG sensors alignment and distortion compen-
sation methodology for 3D scanning applications. Robot. Auton.
Syst. 61(12), 1710–1716 (2013)
123
208
28. Harris, J.W., Stocker, H.: Handbook of Mathematics and compu-
tational Science. Springer, Berlin (1998)
29. He, R., Hu, B.G., Zheng, W.S., Kong, X.W.: Robust principal com-
ponent analysis based on maximum correntropy criterion. IEEE
Trans. Image Process. 20(6), 1485–1494 (2011)
30. Helzer, A., Barzohar, M., Malah, D.: Stable fitting of 2D curves and
3D surfaces by implicit polynomials. IEEE Trans. Pattern Anal.
Mach. Intell. 26(10), 1283–1294 (2004)
31. Huber, P.: Robust Statistics. Wiley, New York (1981)
32. Hunyadi, L., Vajk, I.: Constrained quadratic errors-in-variables fit-
ting. Vis. Comput. 30, 1347–1358 (2014)
33. Karnesky, R.A., Sudbrack, C.K., Seidman, D.N.: Best-fit ellipsoids
of atom-probe tomographic data to study coalescence of γ (L 12 )
precipitates in Ni–Al–Cr. Scr. Mater. 57(4), 353–356 (2007)
34. Keren, D., Cooper, D., Subrahmonia, J.: Describing complicated
objects by implicit polynomials. IEEE Trans. Pattern Anal. Mach.
Intell. 16(1), 38–53 (1994)
35. Keren, D., Gotsman, C.: Fitting curves and surfaces with con-
strained implicit polynomials. IEEE Trans. Pattern Anal. Mach.
Intell. 21(1), 31–41 (1999)
36. Krach, B., Robertson, P.: Integration of foot-mounted inertial
sensors into a bayesian location estimation framework. In: 5th
Workshop on Positioning, Navigation and Communication, 2008
(WPNC 2008), pp. 55–61 (2008)
37. Kubota, T., Okada, K.: Estimating diameters of pulmonary nod-
ules with competition–diffusion and robust ellipsoid fit. In: Liu, Y.,
Jiang, T., Zhang, C. (eds.) Computer Vision for Biomedical Image
Applications, Lecture Notes in Computer Science, vol. 3765, pp.
324–334. Springer, Berlin (2005)
38. Ladrón de Guevara, I., Muñoz, J., Cózar, O.D., Blázquez, E.B.:
Robust fitting of circle arcs. J. Math. Imaging Vis. 40(2), 147–161
(2010)
39. Lee, P.Y., Moore, J.: Geometric optimization for 3D pose estima-
tion of quadratic surfaces. In: Conference Record of the Thirty-
Eighth Asilomar Conference on Signals, Systems and Computers,
2004, vol. 1, pp. 131–135 (2004)
40. Li, Q., Griffiths, J.: Least squares ellipsoid specific fitting. In: Pro-
ceedings of the Geometric Modeling and Processing, 2004, pp.
335–340 (2004)
41. Lou, X., Zhou, L., Jia, Y.: Realization of ellipsoid fitting calibration
for three-axis magnetic sensor based on stm32 embedded system.
In: Zu, Q., Hu, B., Gu, N., Seng, S. (eds.) Human Centered Comput-
ing, Lecture Notes in Computer Science, vol. 8944, pp. 717–726.
Springer, Berlin (2015)
42. Lucas, B.D., Kanade, T.: An iterative image registration technique
with an application to stereo vision. In: Proceedings of the 7th
International Joint Conference on Artificial Intelligence, vol. 2,
pp. 674–679. Morgan Kaufmann, San Francisco, CA, USA (1981)
43. Lukacs, G., Martin, R., Marshall, D.: Faithful least-squares fitting
of spheres, cylinders, cones and tori for reliable segmentation. In:
Burkhardt, H., Neumann, B. (eds.) Computer Vision—ECCV’98,
Lecture Notes in Computer Science, vol. 1406, pp. 671–686.
Springer, Berlin (1998)
44. Mahdavi, S., Salcudean, S.: 3D prostate segmentation based on
ellipsoid fitting, image tapering and warping. In: 30th Annual Inter-
national Conference of the IEEE Engineering in Medicine and
Biology Society, 2008 (EMBS 2008), pp. 2988–2991 (2008)
45. Malyugina, A., Igudesma, K., Chickrin, D.: Least-squares fitting
of a three-dimensional ellipsoid to noisy data. Appl. Math. Sci.
8(149), 7409–7421 (2014)
46. Mathews, V., Xie, Z.: A stochastic gradient adaptive filter with
gradient adaptive step size. IEEE Trans. Signal Process. 41(6),
2075–2087 (1993)
47. Meer, P., Mintz, D., Rosenfeld, A., Kim, D.: Robust regression
methods for computer vision: a review. Int. J. Comput. Vis. 6, 59–
70 (1991)
123
J Math Imaging Vis (2017) 58:189–210
48. Misztal, K., Tabor, J.: Ellipticity and circularity measuring via
Kullback–Leibler divergence. J. Math. Imaging Vis. 55(1),136–
150 (2015)
49. Muñoz-Pérez, J., de Cózar-Macías, O., Blázquez-Parra, E., Ladrón
de Guevara-López, I.: Multicriteria robust fitting of elliptical prim-
itives. J. Math. Imaging Vis. 49(2), 492–509 (2014)
50. Olivares, A., Ruiz-Garcia, G., Olivares, G., Górriz, J.M., Ramirez,
J.: Automatic determination of validity of input data used in ellip-
soid fitting marg calibration algorithms. Sensors 13(9), 11,797
(2013)
51. Oliveira, J., Bioucas-Dias, J., Figueiredo, M.: Adaptive total vari-
ation image deblurring: a majorization–minimization approach.
Signal Process. 89(9), 1683–1693 (2009)
52. Park, S.K., Suh, Y.S.: A zero velocity detection algorithm using
inertial sensors for pedestrian navigation systems. Sensors 10(10),
9163 (2010)
53. Pratt, V.: Direct least-squares fitting of algebraic surfaces. SIG-
GRAPH Comput. Graph. 21(4), 145–152 (1987)
54. Rimon, E., Boyd, S.: Obstacle collision detection using best ellip-
soid fit. J. Intell. Robot. Syst. 18(2), 105–126 (1997)
55. Rousseeuw, P., Leroy, A.: Robust Regression and Outlier Detection.
Wiley, New York (1987)
56. Sahin, T., Unel, M.: Stable algebraic surfaces for 3d object repre-
sentation. J. Math. Imaging Vis. 32(2), 127–137 (2008)
57. Sappa, A., Rouhani, M.: Efficient distance estimation for fitting
implicit quadric surfaces. In: 2009 16th IEEE International Con-
ference on Image Processing (ICIP), pp. 3521–3524 (2009)
58. Sarakon, P., Charoenpong, T., Charoensiriwath, S.: Nose tip detec-
tion using ellipsoid fitting for 2.5D partial face data. In: 2013 6th
Biomedical Engineering International Conference (BMEiCON),
pp. 1–5 (2013)
59. Sarakon, P., Charoenpong, T., Charoensiriwath, S.: Face shape
classification from 3D human data by using SVM. In: 2014 7th
Biomedical Engineering International Conference (BMEiCON),
pp. 1–5 (2014)
60. Shi, J., Tomasi, C.: Good features to track. In: Proceedings of the
IEEE Computer Society Conference on Computer Vision and Pat-
tern Recognition, pp. 593–600 (1994)
61. Skog, I., Nilsson, J.O., Handel, P.: Evaluation of zero-velocity
detectors for foot-mounted inertial navigation systems. In: 2010
International Conference on Indoor Positioning and Indoor Navi-
gation (IPIN), pp. 1–6 (2010)
62. Sturm, P., Gargallo, P.: Conic fitting using the geometric distance.
In: Yagi, Y., Kang, S., Kweon, I., Zha, H. (eds.) Computer Vision—
ACCV 2007, Lecture Notes in Computer Science, vol. 4844, pp.
784–795. Springer, Berlin (2007)
63. Sullivan, S., Sandford, L., Ponce, J.: Using geometric distance fits
for 3D object modeling and recognition. IEEE Trans. Pattern Anal.
Mach. Intell. 16(12), 1183–1196 (1994)
64. Susi, M., Renaudin, V., Lachapelle, G.: Motion mode recogni-
tion and step detection algorithms for mobile phone users. Sensors
13(2), 1539 (2013)
65. Szpak, Z.L., Chojnacki, W., Hengel, A.: Guaranteed ellipse fitting
with a confidence region and an uncertainty measure for centre,
axes, and orientation. J. Math. Imaging Vis. 52(2), 173–199 (2014)
66. Taubin, G.: Estimation of planar curves, surfaces, and nonplanar
space curves defined by implicit equations with applications to edge
and range image segmentation. IEEE Trans. Pattern Anal. Mach.
Intell. 13(11), 1115–1138 (1991)
67. Taubin, G., Cukierman, F., Sullivan, S., Ponce, J., Kriegman, D.:
Parameterized families of polynomials for bounded algebraic curve
and surface fitting. IEEE Trans. Pattern Anal. Mach. Intell. 16(3),
287–303 (1994)
68. Theekapun, C., Hase, H., Tokai, S.: Robust nose localization by
using fitting of ellipse from 2.5d image. In: 2006 IEEE Region 10
Conference on TENCON 2006, pp. 1–4 (2006)
J Math Imaging Vis (2017) 58:189–210
69. Trucco, E., Verri, A.: Introductory Techniques for 3-D Computer
Vision. Prentice-Hall, Upper Saddle River (1998)
70. Vajk, I., Hetthessy, J.: Identification of nonlinear errors-in-variables
models cation of nonlinear errors-in-variables models. Automatica
39, 2099–2107 (2003)
71. Van Den Doel, K., Ascher, U.: The chaotic nature of faster gradient
descent methods. J. Sci. Comput. 51(3), 560–581 (2012)
72. Vandenberghe, L., Boyd, S.: Semidefinite programming. SIAM
Rev. 38(1), 49–95 (1996)
73. Waibel, P., Matthes, J., Gröll, L.: Constrained ellipse fitting with
center on a line. J. Math. Imaging Vis. 53(3), 364–382 (2015)
74. Welsch, R.: Robust regression using iteratively reweighted least-
squares. Commun. Stat. Theory Methods 6(9), 813–827 (1977)
75. Ying, X., Yang, L., Zha, H.: A fast algorithm for multidimensional
ellipsoid-specific fitting by minimizing a new defined vector norm
of residuals using semidefinite programming. IEEE Trans. Pattern
Anal. Mach. Intell. 34(9), 1856–1863 (2012)
76. Yu, J., Kulkarni, S.R., Poor, H.V.: Robust fitting of ellipses and
spheroids. In: Proceedings of the 43rd Asilomar Conference on
Signals, Systems and Computers, pp. 94–98. IEEE Press (2009)
77. Yu, K., Dang, X., Chen, Y.: Robustness of the affine equivariant
scatter estimator based on the spatial rank covariance matrix. Com-
mun. Stat. Theory Methods 44(5), 914–932 (2015)
78. Zhang, Z.: A flexible new technique for camera calibration. IEEE
Trans. Pattern Anal. Mach. Intell. 22(11), 1330–1334 (2000)
79. Zhi-qiang, Z., Guang-Zhong, Y.: Micromagnetometer calibration
for accurate orientation estimation. IEEE Trans. Biomed. Eng.
62(2), 553–560 (2015)
Ezequiel López-Rubio (born
1976) received his M.Sc. and
Ph.D. (honors) degrees in Com-
puter Engineering from the Uni-
versity of Málaga, Spain, in
1999 and 2002, respectively. He
joined the Department of Com-
puter Languages and Computer
Science, University of Málaga,
in 2000, where he is currently
an Associate Professor of Com-
puter Science and Artificial Intel-
ligence. His technical interests
are in machine learning, pattern
recognition and image process-
ing.
Karl Thurnhofer-Hemsi (born
1990) received his B.Sc. in Com-
puter Engineering and his M.Sc.
in Mathematics degrees from the
University of Málaga, Spain, in
2014. He joined the Medical and
Health Research Center of the
University of Málaga in 2015.
He is currently a Ph.D. candidate
at the Department of Computer
Languages and Computer Sci-
ence, University of Málaga. His
technical interests are in medical
image analysis, pattern recogni-
tion and image processing.
209
Óscar David de Cózar-Macías
(born 1972) is associate pro-
fessor with the Department of
Graphic Engineering, Design
and Projects at the University of
Málaga. He is Secretary of the
Polytechnic School from 2001.
He received his P.D. degree from
the University of Málaga in
December 2010, Spain. Also, he
is member of the Graphic Engi-
neering and Design Research
Group and his research inter-
ests are in digital image analysis,
visual pattern recognition, cul-
tural and industrial heritage and product design. He is on the editorial
board of I+Disposed magazine and member of the Spanish Association
of Graphics Engineering.
Elidia Beatriz Blázquez-Parra
(born 1974) is an assistant
professor with the Department
of Graphic Engineering, Design
and Projects at the University of
Málaga. She received her P.D.
degree in Geodesy and Carto-
graphic Engineering in Septem-
ber 2003 from the University
of Jaén, Spain. She is a mem-
ber of the Graphic Engineering
Research Group and her research
interests are in Geodesy, GPS and
geographical information sys-
tems. Her current research activ-
ities include digital image analysis and visual pattern recognition. She is
on the editorial board of I+Disposed magazine and also on the scientific
committee of Topografía y Cartografía journal.
José Muñoz-Pérez (born 1952)
received his M.Sc. in Mathe-
matics from the University of
Granada, Spain, in 1974; and
his Ph.D. degree in Mathemat-
ics from the University of Seville,
Spain, in 1978. He joined the
University of Málaga in 1974,
and then moved to the Univer-
sity of Seville in 1976. Finally he
joined the Department of Com-
puter Languages and Computer
Science, University of Málaga, in
1989, where he is currently a full
Professor of Computer Science
and Artificial Intelligence. His technical interests are in computational
statistics, neural networks and image processing.
123
210 J Math Imaging Vis (2017) 58:189–210

Isidro Ladrón de Guevara-

López (born 1952) received his
P.D. degree in Industrial Engi-
neering in May 1998 from the
University of Málaga. He is head
of the Graphical Engineering
Research Group, and the direc-
tor of the Department of Graphic
Engineering, Design and Project
of Málaga University. He works
on the design process and prod-
ucts development taking out sev-
eral patents, and he is author of
a number of engineering drawing
books. He has published over 30
papers in internationally reputed congress and has participated in differ-
ent Spanish projects of R+D with governments and private companies.
His research activity is also carried out with new researchers and he has
been the advisor of six Ph.D. as well as supervisor of over fifteen. At
last he is on scientific committee of I+Disposed magazine.

123