Fast blending of planar shapes based on invariant
invertible and stable descriptors
1st Emna Ghorbel
CRISTAL Laboratory ( GRIFT group)
Manouba University (National School of Computer Science)
Manouba, Tunisia
e.t.ghorbel@gmail.com
3rd Ines Sakly
CRISTAL Laboratory ( GRIFT group)
Manouba University (National School of Computer Science)
Manouba, Tunisia
adress mail
Abstract—In this paper, a novel method for blending planar
shapes is introduced. This approach is based on the Fined-
Fourier-based Invariant Descriptor (Fined-FID) that is invertible,
invariant under Euclidean transformations and stable. Our
approach extracts the Fined-FID from the two shapes of interest
(the source and the target ones). Then, the extracted descriptors
are averaged enabling the calculation of intermediate descriptors.
Finally, thanks to the inversion criterion, the intermediate shapes
are easily recovered by applying the inverse analytical expression
to these intermediate descriptors. Compared to previous works,
the Fined-FID-based morphing avoid the usual registration step,
generates naturally closed intermediate contours and ensure
invariance under Euclidean transformations and invariance to
the starting point, while being computationally efficient (almost-
linear complexity). The performed experiments show the per-
formance of the proposed blending approach with respect to
curvature-based methods.
Index Terms—Shapes blending, Invariance, Shape morphing.
I. I NTRODUCTION
Shape blending has been an active research topic in the
field of computer graphics and pattern recognition. This is
mainly due to its interest in real-world applications such
as cinematography (face transformation, special effects, etc.),
animation, video games.
Fig. 1. Example of the proposed blending method: the first shape on the left
and the last shape on the right illustrate respectively the source and the target
closed curves, while the shapes in the middle show the generated intermediate
curves.
Shape blending (known also as morphing) aims at building
coherent intermediate shapes between a target shape and a
source one; thus estimating a progressive and continuous
transformation between them. In the literature, a particular
attention has been given to planar shapes [1], [3]–[6] and 3D
2nd Faouzi Ghorbel
CRISTAL Laboratory ( GRIFT group)
Manouba University (National School of Computer Science)
Manouba, Tunisia
faouzi.
4th Slim M’Hiri
CRISTAL Laboratory ( GRIFT group)
Manouba University (National School of Computer Science)
Manouba, Tunisia
adress mail
shapes [1], [10], [11]. In this work, we focus on planar shape
morphing, and more particularly, on the case of closed simple
planar curves. Among the most commonly used planar shape
blending methods, one can mention the linear approach which
consists in linearly interpolating the corresponding points be-
tween two shapes after performing a registration step. Shapira
and Rappoport [4] followed the same idea, but proposed
to construct a Star-Skeleton and interpolated the source and
target skeletons, instead of considering the vertices. In [1],
the authors based their work on the blending of shape interior
polygons [1]. Another way of approaching the problem is to
consider the planar closed shapes as elements of the infinite-
dimensional shape space [5]. The aforementioned approaches
have shown great potential. However, the generated inter-
mediate shapes represent open curves, while they should be
closed. This is mainly due to numerical issues resulting from
the choice of a specific starting point. Furthermore, they are
very sensitive to distortions since they rely on non-continuous
representations. Other blending studies focus on signature
curvature descriptors which are complete and invertible [3],
[6], [12]. These methods consist in interpolating the shape
curvature descriptors. Then, the intermediate descriptors are
used to generate the in-between curves by making use of the
inverse analytical expression of curvature. In order to close the
generated intermediate curves, a closed outline which keeps
the same appearance as the intermediate shapes is optimized.
Among the proposed solutions, the most used one is the local
method B-spline method [7], which affects only a limited
part of the curve. Despite their effectiveness, the curvature-
based methods are computationally expensive due to two main
reasons: (1) the optimization process such as B-spline that
allows closing the intermediate curves and (2) the fact that
they are not naturally invariant to the starting point, relying
therefore on a greedy registration step.
In this paper, we introduce a computationally-efficient
blending method that is by nature invariant to the starting point
and to Euclidean transformations while implicitly ensuring
the generation of intermediate closed curves. In contrast to
previous approaches, our method does not rely on any reg-
istration step or post-processing such as B-spline for closing
the intermediate curves, resulting consequently on an almost-
linearly complexity with respect to the number of curve
points. Our method consists in a generalization of the work
proposed by Surazhsky and al. [6] which is based on curvature
descriptors. However, instead of using curvature descriptors,
we propose to employ the Fourier-Based Invariant Descriptor
(FID) which is complete, invertible, stable and invariant under
the Special Euclidean group denoted by SE(2). For that pur-
pose, a discrete modelization termed fined-FID is introduced.
The advantages presented by the FID and consequently by
the fined-FID are mainly due to the fact that the curve is
described globally thanks to the use of the Fourier coefficients,
in contrast to the local representation of curvatures. Similar to
[6], we define the intermediate curves by linearly interpolating
the Fined-FID functions extracted from the source and the
target shapes and reconstructing the intermediate curves from
the interpolated Fined-FID values using the inverse analytical
expression. We compare our approach to curvature-based ones.
and show that it competes in terms of blending quality and
execution time.
The remainder of this paper is composed of four sections.
Section II formulates the blending problem. In Section III, our
approach is introduced. Section IV depicts the experiments, Fi-
nally, Section V concludes this work and highlights interesting
future directions.
II. P ROBLEM F ORMULATION
Let us denote by C0 and C1 two closed and smooth
curves with an arc-length parameterization (source and target,
respectively). Our goal is to find realistic intermediate curves
between C0 and C1 . More specifically, we aim at finding a
mapping f which returns intermediate curves C0 ∈ M with
M the space of intermediate curves limited by C0 and C1 .
The function f is therefore defined as follows,
f : [0, 1] → M
t 7→ Ct
such that f is continuous, f (0) = C0 and f (1) = C1 .
The curvature based-approach introduced by Surazhsky [6]
is one of the most popular way for blending planar shapes.
Let k1 and k2 be the curvature signature of C1 and C2 ,
respectively. The authors interpolate linearly the two curvature
function in the curvature space to calculate the intermediate
curvature functions as follows,
kt (s) = (1 − t).k1 (s) + t.k2 (s) (1)
with s the arc-length abscissa. Then, they reconstruct the
in-between curve Ct using the inverse curvature expression
k −1 that maps the curvature signatures to the corresponding
Fig. 2. Overview of the Fined-FID based blending approach: the idea is to
go from the shape space to the invariant space, perform linear interpolation in
the invariant space, then return to the shape space to obtain the intermediate
curves.
point shapes in the shape space. In other words, for t ∈ [0, 1],
an in-between curve f (t) is computed as follows,
f (t) = k −1 ((1 − t).k(C1 ) + t.k(C2 )). (2)
The effectiveness of this approach is mainly due to the
fact that the curvature descriptor verifies the four necessary
properties described below.
(1) Completeness: a complete descriptor means that two
objects have the same shape if and only if they have the same
set of descriptors [9]. This criterion is essential since it ensures
a one-to-one correspondence between the shape space and the
curvature signature space
(2) Invertibility: an invertible descriptor means that the
reconstruction of objects from their description is possible
through an explicit analytical formula. In [6], this enables the
reconstruction of intermediate curves from the intermediate
interpolated curvature signatures.
(3) Invariance under SE(2): this means that two shapes
that are similar up to a Euclidean transformation should have
the same set of descriptor. This makes the blending approach
robust to Euclidean transformations.
(4) Stability: a stable descriptor means that a low level
distortion of the shape does not induce a noticeable change in
the set of descriptor [8]. This makes the shape blending robust
to small distortion. In [6], this results from the continuity of
the curvature function.
Although the curvature signature have shown great per-
formance in the context of shape blending, there exist still
some limitations. First, open in-between curves are usually
generated. This problem is due to the numerical successive
derivations and the non-periodicity of the descriptor which is
local. In other words, the starting point and the latest point
of the curve are numerically not confused. Second, a pre-
processing step of registration is needed to align the two
starting points of the source and target shapes; thus, increasing
the computational complexity.
 The aim of this paper is to overcome these two issues while
taking advantage of the strong theoretical background resulting
from the definition of the four aforementioned criteria. There-
fore, the core idea consists in generalizing the curvature-based
approach proposed in [6] by selecting or/and designing a more
adequate descriptor. This descriptor should not only respect the
four properties defined earlier, but also verify two additional
properties described below:
(1) A global description power: Unlike the curvature signa-
ture, the chosen descriptor should describe the shape globally Fig. 3. Steps of the Fined-FID based blending: a target curve and a source
and not in each point of the shape. This will allow the curve are given as input. The Fined-FID descriptor is then extracted from
the two curves. Then, the intermediate fined-FID are computed by averaging
generation of closed intermediate curves. the target and source fined-FID descriptors. Finally, the inverse Fined-FID
(2) Invariance to the starting point: Two shapes that are formula is applied allowing the computation of intermediate curves in the
similar with different starting point should naturally have the shape space.
same set of descriptor. This excludes the use of any registration
pre-processing.
with t ∈ [0, 1]. Figure 3 illustrates the different steps under-
This can be summarized as follows: The intermediate shapes
taken for carrying out the shape blending using the Fined-
f (t) between C1 and C2 with t ∈ [0, 1] will be generated
FID. In the following, we start by recalling the FID. Then, the
using a global descriptor g that is complete, invertible, stable,
Fined-FID is introduced while taking into account numerical
invariant to SE(2) and to the starting point as described below,
aspects.
f (t) = g −1 ((1 − t).g(C1 ) + t.g(C2 )) (3) A. Fourier-Based Invariant descriptor
In next section, we discuss the choice of g that is essential In this part, we recall the FID [8] formula as well its inverse
for ensuring the effectiveness of the described descriptor-based expression.
blending approach. Let p, q be two non-zero positive reals, n0 , n1 be two
integers (n0 > n1 ) such that they represent the index of the
III. F ORMULATION OF THE BLENDING APPROACH BASED
Fourier coefficients with the highest energy, s be a periodic
ON THE F INED -F OURIER I NVARIANT DESCRIPTOR
normalised arc-length parameterization of a closed planar
For the shape blending, we choose the Fourier-based In- curve C and an be the FourierPcoefficients of C [8]. Given
variant Descriptor (FID) [8] which verifies all the necessary that l2 (Z) = {(xn )n∈Z ∈ CZ ; n∈Z |xn |2 < ∞} and CZ is
properties. the set of functions from Z to C, If {an } is in l2 (Z) with n
In the literature, several approaches made use of the Fourier the number of Fourier coefficients then the sequence In which
Descriptor or proposed an extension of the latter ( [14]–[18]). is the FID function is calculated as follows,
Different variants of the Fourier descriptor have been used for
numerous tasks such as pattern recognition and compression.  n0 −n1 n1 −n n−n0
an an an
However, applying these descriptors to the context of shape In
 = 0 1
, ∀n 6= n0 ,n1
|an0 |n1 −n−p |an1 |n−n0 −q

blending is not straightforward. This is principally caused by In0 = |an0 |n0 −n1 +p
, n = n0 and an0 6= 0
their lack of stability. 
|an1 |n1 −n0 +q , n = n1

As mentioned earlier, unlike previous Fourier-based descrip-
 I n1 = 6= n0
an1 6= 0
and
(5)
tors approaches, FID [8] verifies, in addition the completeness,
invariance to Euclidean transformations and invertibility, the we precise that we choose p and q near to zero in the spirit of
fourth criterion of stability. Moreover, FID is a global and peri- creating an isometry between the shape space and the invariant
odic description, verifying, therefore the two additional criteria space. An optimization study should be carried out to choose
defined in this paper. This enables avoiding the commonly the best p and q.
used registration step and solving the closing curve problem. To calculate the FID descriptor inverse formula, Equation 5
For that reason, we consider that this descriptor is an ideal has been inverted. This gives,
candidate for our generalized blending method.
−p −q
Moreover, we propose a discrete and fined modelization
ann0 −n1 = In In 0n0 −n1 +p Inn11 −n0 +q e2iπ(n−n1 )θ0 e2iπ(n0 −n)θ1 ,
of FID termed Fined-FID. The computation of Fined-FID is
(6)
based on DFT coefficients.
Thereafter, it results,
Figure 2 describes the proposed blending method which is
based on the Fined-FID. 1 −p −q
2iπ
Let assume I to be the Fined-FID and I −1 its reverse an = Inm0 Inα01 Inα12 e m0 (n(θ0 −θ1 )−n1 θ0 +n0 θ1 ) , (7)
expression. Then, Equation 3 becomes,
such that m0 = n0 − n1 , α1 = m0 (n0 − n1 + p), α2 =
f (t) = I −1 ((1 − t).I(C0 ) + t.I(C1 )), (4) m0 (n1 − n0 + q).θ0 and θ1 are the arguments of an0 and an1
respectively. They represent the two Fourier coefficients with
the maximum energy of the parameterized contour C.
Subsequently, the corresponding curve can be recovered by
simply calculating the Fourier series of the coefficients an .
B. Fined-FID and numerical considerations
In this subsection, we introduce the Fined-FID which is
the FID associated discrete modeling. The idea is to replace
the Fourier Transform coefficients by the Discrete Fourier
Transform (DFT) coefficients . Then, we study the numerical
processing of the FID-Fined descriptor. Finally, we report the
complexity of the Fined-FID based blending algorithm.
1) Discrete Fourier Transform: Concretely, the curve that
are processed are fined and discrete. In practice, we have a
finite number of periodic samples. Thus, adapting FID to a
discrete space is required. For that purpose, we choose to
replace the Fourier coefficient by the Discrete Fourier coeffi-
cients (DFT) described below. For a closed planar curve after
an arc-length parameterization, we obtain zk = xk + iyk ∈ C
such that k ∈ {0, 1, .., N − 1}, the DFT coefficient is given
by,
3) Truncation: In order to apply our approach which is
based on Fined-FID method, we consider the truncation step
which is necessary for avoiding numerical complications. The
idea behind the truncation is to consider only a window around
the highest DFT energy coefficients and remove the energy
outside the window to provide an effective algorithm. This
technique, therefore, avoid the inclusion of infinite values,
negligible details or noise in the description function. Indeed,
the powers of the DFT coefficient indices are included in
the Fined-FID formula. Thus, this can make the fined-FID
divergent if coefficients with high indices are taking into
consideration.
A practical example of truncation is shown in Figure 4
where the chosen window corresponds to the interval [n0 −
N/10, n0 + N/10] with N = 150 the number of shape
coefficients and n0 the highest DFT energy coefficient index.
Then, it becomes possible to calculate the inverse Fined-FID
formula from the truncated DFT without facing any numerical
issues.

N −1
1 X kn
Fn = zk e−2iπ N (8)
N
k=0

Similarly, the Fourier series is replaced by the Inverse

Discrete Fourier Transform IDFT defined as follows, Fig. 4. The signal A represents the descriptor coefficient module after the
application of the DFT on the shape C. The signal B represents the module
N −1
of the descriptor coefficients after the application of the DFT on the same
X kn shape C with truncation
zk = Fn e2iπ N (9)
n=0

2) Fined-Fourier-based Invariant Descriptor: Let p, q be

two non-zero positive reals, n0 , n1 be two integers (n0 > n1 ),
s be a periodic normalised arc-length parameterization of a
closed planar curve C and Fn be the DFT coefficients of the
curve. The sequence Jn is calculated as In (Equation 5) using
DFT coefficients, as depicted below,


Fnn0 −n1 Fnn1 −n Fnn−n0
Jn

 = 0 1
|Fn0 |n1 −n−p |Fn1 |n−n0 −q
, ∀n 6= n0 ,n1
n0 −n1 +p
 Jn0 = |Fn0 | , n = n0 and Fn0 6= 0
|Fn1 |n1 −n0 +q , n = n1

 J n1 = 6= n0
Fn1 6= 0 and
(10)
The Fined-FID inverse formula is also similar to the inverse
FID formula (Equation 7), but incorporates DFT coefficients Fig. 5. Truncation of the original curve with various percentages. Note that
as follows, the higher the percentage of truncation is, the less detailed the reconstructed
curve is.
1 −p −q
2iπ
Fn = J n Jm0 α1
n0
α2
J
n1 e m0 (n(θ0 −θ1 )−n1 θ0 +n0 θ1 ) (11) We remark in Figure 5 that the truncation can be seen as
a filter which makes the shapes smoother. In other words, the
with θ0 and θ1 the respective arguments of Fn0 and Fn1 . higher the percentage of the truncation is the more smoothed
After that, starting from Fn which represents the DFT the shape is (until becoming unknowable). An optimization
coefficients, an IDFT is applied to rebuild the initial curve study should be done on the choice of the truncation percent-
C. age depending on the shape.
 4) Complexity: To calculate the Fined-FID morphing algo-
rithm complexity, we assume that the planar closed shape has
already an arc-length parameterization.
Given that the matching step is avoided, we observe that the
complexity of all the operations except for the Discrete Fourier
Transform is equal to O(n) with n the number of curve points.
The Discrete Fourier Transform requires a complexity of
O(n2 ). However, it can be easily replaced by the Fast Fourier
Transform, resulting on an important decrease of the com-
plexity to O(n.log(n)). Hence, the algorithm complexity is
almost-linear in contrast to other methods where a registration
step is necessary and leads to an increase of the complexity. Fig. 6. Transformation of the source curve C0 (on the left) to the target
The registration methods necessitates usually a complexity of curve C1 (on the right). The first line shows the linear interpolation based on
O(n2 ). Furthermore, in curvature-based approaches such as the parameterization, the second line the interpolation based on the curvature,
the third line the results of the second line with a b-spline closure method
[3], [6] , a b-spline method is generally used to close the and the fourth line shows our method which is an interpolation based on the
open in-between curves. Thereafter, knowing that b-splines descriptor Fined-FID. The spot indicates the beginning and the end of the
algorithm is of O(n2 ) complexity, the curvature blending have curves.
a higher complexity than the Fined-FID blending.
to naturally close the curves without the need of any post-
IV. E XPERIMENTS
processing such as B-Splines, in contrast to curvature-based
In this section, we report and analyze the experimen- approaches. This is also confirmed by the qualitative results
tal results. First, we compare the proposed approach based shown in Figure 6.
on the Fined-FID descriptor to the simple curvature based
method, then to curvature-based methods with b-splines [3],
[6]. Finally, qualitative results are reported on the KIMIA99
dataset which incorporates multiple closed curves with dif-
ferent shapes. Intra-class as well as inter-class intermediate
shapes are generated using this dataset.
To blend shapes with our method, we follow these steps:
(1) We extract the Fined-FID descriptor from both the source
shape C and the target one C 0 . (2) After truncation, we
interpolate linearly the 2D discrete signal source and the
target one to get the weighted average signals. (3) Finally, we
reconstruct the intermediate shapes using the inverse analytical
expression. Figure 3 summarizes these different steps.
Fig. 7. Transformation of the same source and target curves except that
A. Comparison with curvature methods we apply a 3π/2 rotation on the curve C1 . The first line shows the linear
interpolation based on the parameterization. The second one illustrates the
In this subsection, we compare our method to curvature- interpolation based on the curvature. The third line presents the results of the
second line with a b-spline closure method. Finally, the fourth line shows our
based approaches. Different aspects are mentioned, namely, method. We note that the rotation did not affect the intermediate curves of the
the morphing quality, the invariance to Euclidean transforma- methods that are based on invariant descriptors (Curvature and Fined-FID)
tions and the invariance to the starting point.
1) Morphing quality: In Figure 6, we show a deformation 2) Invariance to Euclidean transformations: To evaluate
of the curve C0 towards the contour C1 and then compare the robustness to Euclidean transformations, we propose to
the intermediate curves constructed using a blending method apply an indirect rotation of 90° to the target contour C1 .
based on the linear interpolation of parameterization, the The results are illustrated in Figure 7. We observe that the
one based on the curvature, the one based on the curvature intermediate curves of the proposed method is invariant to
with b-spline [3], [6] to our approach (Fined-FID blending Euclidean transformations. This means that the generated
method). The results suggest the superiority of our approach in intermediate curves are conserved up to a Euclidean trans-
terms of quality. In fact, the evolution of intermediate shapes formation, in contrast to the parameterization-based approach.
generated using our approach is visually smoother than the This results mainly from the invariance property of the Fined-
ones computed using previous approaches. In fact, while our FID descriptor and the curvature-based descriptor.
approach modifies the shape progressively while conserving 3) Invariance to the starting point: It is important to note
the concavities, the other approaches tend to delete them, then that the starting point for the Fined-FID method was chosen
create new ones for reaching the target shape. furthermore, randomly, in contrast to the two curvature-based methods.
thanks of its global description power, our approach is able Thus, we observe in Figure 8 the invariance of the generated
Fig. 8. Transformation of the source curve C0 (on the left) to the target curve
C1 (on the right). The first line and the second line show the application of
the proposed method with two different starting points

Fig. 9. Transformation of the source curve C0 (on the left) to the target curve
C1 (on the right). The first line and the second line show the application of
the curvature-based blending method with two different starting points

intermediate curves using our approach despite the change of

the starting point. In Figure 9, we clearly show the difference
of the two deformations generated by the same contours using
the curvature-based approach. This shows that the intermediate
curves generated by this approach strongly depend on the
choice of the starting point.
B. Morphing on Kimia99 Data base
We apply our method on the KIMIA99 dataset. We observe
in Figure 10 and Figure 11 the morphing of these contours
using the Fined-FID morphing. It can be remarked that the
qualitative results are good in both cases: (1) if the target
and source shapes belong to the same class and (2) if the
two shapes belong to different classes. While the first case
has been widely studied, the second one is not obvious,
since usually a first step of registration is necessary. In fact, Fig. 10. Application of our morphing method on KIMIA99 database
aligning two shapes belonging to two different classes is not
straightforward. Therefore, since it is completely independent
of any matching step, our approach is able to generate mean- of the invariant space is not known, its simple connectivity
ingful intermediate shapes between two shapes of different is not guaranteed. Hence, we are planning to carry out an
classes. This advantage results mainly from the globality of optimization study to find the (non-linear) path in the invariant
the Discrete Fourier Coefficients which are the basis of the space. This might lead to the generation of more relevant
Fined-FID construction. intermediate curves.

V. C ONCLUSION
In this article, we introduce a general approach for the
interpolation of continuous and closed curves based on invari-
ant descriptors. For that purpose, the Fined-FID is introduced
which is a discrete modelization of the FID. We prove that
our approach is invariant to Euclidean transformations and
invariant to the starting point. Thus, the registration steps
commonly applied between the target and source curves are
avoided. This allows a considerable reduction of the computa-
tional complexity. In future works, we aim at considering the
interpolation of open curves. Indeed, the current approach that
is based on DFT coefficients is not applicable if the source and
the target curves are open. Furthermore, since the topology
R EFERENCES
[1] M. Alexa, D. Cohen-Or, and D. Levin. As-rigid-as-possible shape
interpolation. InProceedings ofthe 27th annual conference on
Computer graphics and interactive techniques, pages 157–164.
ACMPress/Addison-Wesley Publishing Co., 2000.
[2] F. Ghorbel.Invariants de formes et de mouvement 11 cas du 1D au 4D
et de l’euclidien aux projectifs.La Manouba : ARTS-PI edition., 2013.
[3] M. Saba, T. Schneider, K. Hormann, and R. Scateni. Curvature-based
blending of closed planar curves.Graphical Models, 76(5):263–272,
2014.
[4] M. Shapira and A. Rappoport. Shape blending using the star-skeleton
representation.IEEE ComputerGraphics and Applications, 15(2):44–50,
1995.
[5] KLASSEN, Eric, SRIVASTAVA, Anuj, MIO, M., et al. Analysis of
planar shapes using geodesic paths on shape spaces. IEEE transactions
on pattern analysis and machine intelligence, 2004, vol. 26, no 3, p.
372-383.
 of Passiflora leaves: the relationship between landmarks of the vascula-
ture and elliptical Fourier descriptors of the blade. GigaScience, 2017,
vol. 6, no 1, p. giw008.
[16] DE BAETS, Leen, DEVELDER, Chris, DHAENE, Tom, et al. Auto-
mated classification of appliances using elliptical fourier descriptors. In
: 2017 IEEE International Conference on Smart Grid Communications
(SmartGridComm). IEEE, 2017. p. 153-158.
[17] BENITEZ-GARCIA, Gibran, NAKAMURA, Tomoaki, et KANEKO,
Masahide. Facial expression recognition based on local fourier coef-
ficients and facial fourier descriptors. Journal of Signal and Information
Processing, 2017, vol. 8, no 3, p. 132-151.
[18] LIN, C. C. et CHELLAPPA, Rama. Classification of partial 2-D shapes
using Fourier descriptors. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 1987, no 5, p. 686-690.

Fig. 11. Application of the Fined-FID morphing on KIMIA99 database

[6] T. Surazhsky and G. Elber. Metamorphosis of planar parametric curves

via curvature interpolation.International Journal of Shape Modeling,
8(02):201–216, 2002.
[7] P. Thevenaz, T. Blu, and M. Unser. Interpolation revisited [med-
ical images application].IEEE Trans-actions on medical imaging,
19(7):739–758, 2000.
[8] GHORBEL, Faouzi. Stability of invariant Fourier descriptors and its
inference in the shape classification. In : 11th IAPR International
Conference on Pattern Recognition. Vol. III. Conference C: Image,
Speech and Signal Analysis,. IEEE, 1992. p. 130-133.
[9] CRIMMINS, Thomas R. A complete set of Fourier descriptors for
two-dimensional shapes. IEEE Transactions on Systems, Man, and
Cybernetics, 1982, vol. 12, no 6, p. 848-855.
[10] ALHASHIM, Ibraheem, LI, Honghua, XU, Kai, et al. Topology-varying
3D shape creation via structural blending. ACM Transactions on Graph-
ics (TOG), 2014, vol. 33, no 4, p. 1-10.
[11] WU, Bo, XU, Kai, ZHOU, Yang, et al. Skeleton-guided 3D shape
distance field metamorphosis. Graphical Models, 2016, vol. 85, p. 37-45.
[12] HIRANO, Masahiro, WATANABE, Yoshihiro, et ISHIKAWA,
Masatoshi. Rapid blending of closed curves based on curvature flow.
Computer Aided Geometric Design, 2017, vol. 52, p. 217-230.
[13] ZAHN, Charles T. et ROSKIES, Ralph Z. Fourier descriptors for plane
closed curves. IEEE Transactions on computers, 1972, vol. 100, no 3,
p. 269-281.
[14] PERSOON, Eric et FU, King-Sun. Shape discrimination using Fourier
descriptors. IEEE Transactions on systems, man, and cybernetics, 1977,
vol. 7, no 3, p. 170-179.
[15] CHITWOOD, Daniel H. et OTONI, Wagner C. Morphometric analysis