Page 1
3.5 Comparison Results of Nonlinear Dimensionality Reduction Techniques 25
0 5 10 15 20
-10
-5
0
5
10
(a) PCA
-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05 0.06
-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
(b) ISOMAP
-0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
(c) LLE
(d) LEM
-20 -15 -10 -5 0 5 10 15
-6
-4
-2
0
2
4
6
(e) MVU (f) t-SNE
Figure 3-6.: Nonlinear Feature Extraction the Swiss roll with hole database
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
(a) PCA
-0.04 -0.02 0 0.02 0.04 0.06
-0.06
-0.04
-0.02
0
0.02
0.04
(b) ISOMAP
-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
(c) LLE
(d) LEM
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
(e) MVU (f) t-SNE
Figure 3-7.: Nonlinear Feature Extraction for the fishbowl database
Page 2
26 3 Comparative Analysis of Nonlinear Feature Extraction Techniques
3-3 and Table 3-4, present the quality of the embedding for the real-world datasets. On the
other hand, Table 3-5 and 3-6 show the computational time for all the methods for 2-D and
3-D representations.
Page 3
3.5 Comparison Results of Nonlinear Dimensionality Reduction Techniques 27
1.2 1.25 1.3 1.35 1.4 1.45
x 109
-6
-4
-2
0
2
4
6
x 107
(a) PCA (b) ISOMAP
(c) LLE
-0.06 -0.04 -0.02 0 0.02 0.04 0.06
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
(d) LEM
-200 -150 -100 -50 0 50 100 150 200
-150
-100
-50
0
50
100
150
(e) MVU (f) t-SNE
(g)
Figure 3-8.: 2-D representation for the Manki Neko database
Page 4
28 3 Comparative Analysis of Nonlinear Feature Extraction Techniques
4 4.5 5 5.5 6
x 108
-2.5
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
x 107
(a) PCA
-0.15 -0.1 -0.05 0 0.05 0.1 0.15
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
(b) ISOMAP
(c) LLE
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
(d) LEM
-150 -100 -50 0 50 100 150
-100
-50
0
50
100
(e) MVU (f) t-SNE
(g)
Figure 3-9.: 2-D representation for the Frog database
Page 5
3.5 Comparison Results of Nonlinear Dimensionality Reduction Techniques 29
(a) PCA (b) ISOMAP
(c) LLE
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.05
-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
0.04
0.05
(d) LEM
-150 -100 -50 0 50 100 150
-150
-100
-50
0
50
100
150
(e) MVU (f) t-SNE
(g)
Figure 3-10.: 3-D representation for the Duck database
Page 6
30 3 Comparative Analysis of Nonlinear Feature Extraction Techniques
(a) PCA (b) ISOMAP
-0.1
0
0.1
-0.15-0.1
-0.050
0.050.1
0.15
-0.1
0
0.1
(c) LLE
-0.5
0
0.5
1
-1-0.8
-0.6-0.4
-0.20
0
0.5
1
(d) LEM
-200
-100
0
100
200-150
-100-50
050
100150
-20
0
20
(e) MVU (f) t-SNE
(g)
Figure 3-11.: 3-D representation for the Manki Neko database
Page 7
3.5 Comparison Results of Nonlinear Dimensionality Reduction Techniques 31
(a) PCA (b) ISOMAP
(c) LLE
-0.4
-0.2
0
0.2
0.4
0.6
0.8
-1
-0.5
0
00.5
1
(d) LEM
-150
-100
-50
0
50
100
150-100
-500
50100
-20
0
20
(e) MVU (f) t-SNE
(g)
Figure 3-12.: 3-D representation for the Frog database
Page 8
32 3 Comparative Analysis of Nonlinear Feature Extraction Techniques
(a) PCA (b) ISOMAP
(c) LLE
-1
-0.5
0
0.5
-1
-0.5
0
0.5
-0.50
0.51
(d) LEM
-150
-100
-50
0
50
100
150
-150-100
-500
50100
150
-200
20
(e) MVU (f) t-SNE
(g)
Figure 3-13.: 3-D representation for the Duck database
Page 9
3.5 Comparison Results of Nonlinear Dimensionality Reduction Techniques 33
Table 3-1.: PNE for synthetic datasets, ideal PNE= 0
Method Database
Fishbowl S-Surface Swiss Role with Hole
ISOMAP 0, 4393 0, 3660 0, 5135
LLE 0, 0062 0,0009 0, 1020
LEM 0,0032 0, 0052 0, 238
MVU 0, 0121 0,0009 0,0031
Table 3-2.: Computational load in seconds. 2-D representations of the synthetic datasets.
Method Database
Fishbowl S-Surface Swiss Role with Hole
ISOMAP 52, 38 53, 05 49, 74
LLE 24, 25 23, 67 26, 30
LEM 81, 57 83, 11 33, 35
MVU 13.8987 8.178, 2 9.592, 4
PCA 0,07 0,01 0,01
t-SNE 83, 91 87, 01 96, 13
Table 3-3.: PNE for real-world datasets – 2-D representations, ideal PNE = 0
Method Database
Maneki Neko Frog Duck
ISOMAP 0, 0427 0, 0686 0, 0575
LLE 0, 0444 0, 0404 0,0221
LEM 0,0184 0,0167 0, 0334
MVU 0, 0492 0, 0538 0, 0342
Table 3-4.: PNE for real-world datasets – 3-D representations, ideal PNE = 0
Method Database
Maneki Neko Frog Duck
ISOMAP 0, 2085 0, 1939 0, 2436
LLE 0, 0418 0, 0418 0,0205
LEM 0,0220 0,0276 0, 0483
MVU 0, 0454 0, 0451 0, 0310
Page 10
34 3 Comparative Analysis of Nonlinear Feature Extraction Techniques
Table 3-5.: Computational load in seconds. 2-D representations of the real-world datasets.
Method Database
Maneki Neko Frog Duck
ISOMAP 1, 90 1.95 1, 96
LLE 2, 25 2, 27 2, 34
LEM 3, 05 2, 59 2, 23
MVU 3, 39 3, 74 3, 88
PCA 0,03 0,31 0,04
t-SNE 1, 02 1, 38 0, 95
Table 3-6.: Computational load in seconds. 3-D representations of the real-world datasets.
Method Database
Maneki Neko Frog Duck
ISOMAP 1, 90 1, 93 1, 87
LLE 2, 23 3, 10 2, 36
LEM 2, 33 2, 42 2, 37
MVU 3, 30 3, 69 3, 87
PCA 0,03 0,03 0,03
t-SNE 1, 08 1, 00 0, 97
Page 11
3.6 Discussion 35
3.6. Discussion
The conventional PCA does not achieve adequate results for the synthetic datasets, due to
the low-dimensional representations exhibit overlapping, as can be seen in Figures 3.7(a),
3.5(a) and 3.6(a). That is, PCA does not seek to unfold the input data, but it finds the
projection that preserves the variance of data as much as possible without taking into account
neither the local nor global data structure. Similar results are obtained for the real-world
datasets, where the behavior of the activity is lost in the mapping process(Figures 3.8(a),
3.9(a), 3.10(a), 3.11(a), 3.12(a) and 3.13(a)). In this regard, PCA is not appropriate for
analyzing data that lie in nonlinear manifold, specially for visualization tasks.
In contrast, other methods like ISOMAP, LLE, LEM and MVU find low-dimensional
spaces that suitably represent the data. However, the performance of these techniques is
not constant for all the databases. For instance, ISOMAP is not able to properly unfold the
Fishbowl dataset (see Figure 3.7(b)), but shows better results for the S-Surface and the Swiss
Role with Hole, getting faithful representations of the high dimensional data (Figures 3.5(b)
and 3.6(b)). It is possible to observe that in these cases some holes or blank patches appear
in the output space, which is one of the weakness of ISOMAP that is common when data is
not uniform and well sampled. For the COIL-100 database, the 2-D embeddings obtained
with ISOMAP exhibit a symmetrical and consistent global structure (Figures 5.11(a), 3.9(b)
and 3.10(b)), specially for the Maneki Neko and Duck datasets. For The 3-D embeddings,
ISOMAP tries to preserve the global data structure, but the results are not as smooth as
the other tested methods (Figures 3.11(b), 3.12(b) and 3.13(b)).
LLE demonstrates a good performance dealing with synthetic data. Nevertheless, there
is a great influence of the regularization parameter (when k > p) in low-dimensional repre-
sentations. According to the results (Figures 3.7(c), 3.5(c) and 3.6(c)), this method seeks
to unfold input data preserving both global and local structure, but most of the times it
produces a change of scale or rotation in the low-dimensional space, as in the Swiss Role
with Hole and S-Surface datasets. Table 3-1 shows that the quality of LLE embeddings are
not the lower ones, and some times it gets the best transformation. Again, for the real-world
datasets, LLE finds suitable low-dimensional representations that clearly allow to identify
the object rotation, which is due to the preservation of local properties of the input data,
as shown in Figures 3.8(c), 3.9(c) and 3.10(c). In this sense, the results are quite similar to
those found with other techniques such as LEM and ISOMAP for the 2-D representations.
Now, for the 3-D embeddings the performance of LLE is higher than ISOMAP according to
Table 3-4, which can be visually confirmed in Figures 3.11(c), 3.12(c) and 3.13(c).
The LEM method has difficulties to compute appropriate transformations for the artificial
datasets. Even when it tries to unfold data in order to conserve the global properties, the
local relationships are not fully preserved, which generates two possible consequences: blank
patches or overlapping in the low-dimensional space. These characteristics are evident in
Figures 3.7(d), 3.5(d) and 3.6(d). This may be explain by considering that LEM attempts
Page 12
36 3 Comparative Analysis of Nonlinear Feature Extraction Techniques
to be insensitive to outliers, and for that reason if the manifold is not well sample or is
non-uniform it adds a penalty and the local geometry is not properly conserve. Nonetheless,
LEM has the higher performance for the Fishbowl dataset, which is one of the most difficult
to unfold considering its intrinsic geometry. In general, it is possible to say that LEM
is not the best option to handle synthetic datasets for visualization purposes. Otherwise,
the embeddings obtained for the real-world databases are much better, and LEM shows the
higher performance in two of the three tested databases for both 2-D and 3-D representations
(see Table 3-3 and 3-4). In this sense, the LEM technique proves to be a good option to
work with real-word data, finding embeddings that preserve the local and global structure.
Moreover, it has a good performance for both 2-D and 3-D visualization tasks (Figures
3.8(d), 3.9(d), 3.10(d), 3.11(d), 3.12(d) and 3.13(d)).
In regard to the synthetic datasets, MVU presents the higher performance for the S-surface
and the Swiss Roll with Hole, as can be confirm in Table 3-1. This technique seeks to find a
transformation that preserves both global and local data structure simultaneously, which is
implicit in the constrains of the optimization problem. As a consequence, data is properly
unfolded in the low-dimensional space, exhibiting symmetry (see Figures 3.5(e) and 3.6(e)).
Specifically, for the Fishbowl, MVU tries to unfold the data points but the output space
shows overlapping 3.7(e). It is important to note that this manifold has regions with low
density of samples, which affects the performance of the algorithms. MVU’s weakness lies
in the fact that the SDP optimization problem requires a high computational load, and for
that reason while other techniques only need a few minutes to perform the analysis, MVU
demands hours, see Table 3-2. This time demand decreases when the amount of samples is
lower, and therefore the computational load for the tested real-world datasets is acceptable
and competitive with the other techniques (see Table 3-5 and 3-6). However, the quality of
the embedding results is not as good for the real-world datasets. In this cases, the results still
show symmetry, but the local data structure is not well preserved. So, even if the rotation
behavior can be infer from the low-dimensional representation, the local properties are not
present in the output space (Figures 3.8(e), 3.9(e), 3.10(e), 3.11(e), 3.12(e) and 3.13(e)).
Finally, with the purpose of comparing the manifold learning methods with another kind
of techniques for dimensionality reduction, the t-SNE algorithm was taken into account in
this analysis. Unlike the above techniques, t-SNE is based on probabilistic theory, which add
another type of parameters to the transformation related to the chosen distribution. With
respect to the artificial databases the low-dimensional representations found with t-SNE
have a common characteristic, they present a considerable amount of blank patches, thence
the data seems to be disconnected, and the local properties of the input data are not fully
conserved,, as can be notice in Figures 3.7(f), 3.5(f) and 3.6(f). Another weakness about the
t-SNE method is to tune the initial condition of the algorithm, which can change every time
that the method is used, affecting the transformation, and hence, the output is quite different
every time. For the real-world datasets the results are similar, the embeddings do not reflect
the rotation behavior, and the local structure of data is not preserved by neither the 2-D
Page 13
3.6 Discussion 37
nor 3-D representations (see Figures 3.8(f), 3.9(f), 3.10(f), 3.11(f), 3.12(f) and 3.13(f)).
Page 14
4. Multiple Manifold Learning for
Dimensionality Reduction
Often, in machine learning and pattern recognition literature, the nonlinear feature extrac-
tion (NFE) techniques are reviewed as learning methods for discovering an underlying low-
dimensional structure from a set of high-dimensional input samples, that is, NFE techniques
unfold a non-linear manifold embedded within a higher-dimensional space. Nevertheless,
most of the NFE algorithms are constrained to deal with a single manifold, attaining unap-
propriate low-dimensional representations when input data lie on multiple manifold, because
the inter-manifold distance is usually much larger than the intra-manifold distance [33], mov-
ing apart each manifold from the others, regardless of whether the behavior among them is
similar.
To our best knowledge, some few works [23, 33, 34] have been proposed in order to allow
the application of the NFE techniques on multiple manifold data sets. Particularly, in [33]
a framework to learn an embedded manifold representation from multiple data sets called
Learning a Joint Manifold (LJM) is presented, which finds a common manifold among the
different data sets, without assuming some kind of correspondence between the different
manifolds. However, the main drawback of this approach is that for obtaining suitable
low dimensional representations the input samples must be similar in appearance, when
the multiple manifolds do not have a close resemblance among them the LJM method fails
to embed the data. On the other hand, the approach presented in [34] actually uses a
correspondence labels among the samples in order align the data sets, in such case the
complexity of the challenge is lower than when no one correspondence is assumed. A similar
solution is proposed in [35].
Unlike these mentioned works for dealing with multiple manifolds, our major contribution
to the state of the art is the possibility for analyzing dissimilar objects/subjects in appearance
but with a common behavior (similar motion), moreover our methodology allows to employ
objects/subjects with different input dimensions and number of samples among manifolds.
Our approach is inspired on the manifold learning algorithm Laplacian Eigenmaps - LEM
[18](described in Section 3.1.5), because its optimization problem has an analytic solution
avoiding local minima, and few free parameters need to be fixed by user. Our approach can
be employed to visually identify in a low-dimensional space the dynamics of a given activity
learning it from a variety of datasets. We test on two real-world databases, changing the
number of samples and input dimensions per manifold. Our proposal is compared against
Page 15
4.1 Multiple Manifold Learning - MML 39
both the conventional Laplacian Eigenmaps (LEM) [23] and the closest work found in the
state of the art for multiple-manifold learning (LJM) [33]. Overall, our methodology achieves
meaningful low dimensional representations, visually outperforming the results obtained by
the other methods.
Next section (4.1) introduces the proposed methodology for multiple manifold dimension-
ality reduction. Then, the experimental conditions are presented, and finally the results are
described and discussed.
4.1. Multiple Manifold Learning - MML
The NFE frameworks based on manifold learning such as ISOMAP, LLE and LEM were
shown to be able to embed data lying on a nonlinear manifold into a low-dimensional Eu-
clidean space for synthetic datasets, as well as for real images (for example the COIL-100
Database tested in Chapter 3). Nonetheless, these techniques fail when they look for a com-
mon low-dimensional representation for data lying on multiple manifolds or multiple classes.
Considering that in real world applications, like video and image analysis, datasets with mul-
tiple manifolds are common, we propose a framework to find a low-dimensional embedding
for data lying on multiple manifolds. The proposed approach is inspired on the manifold
learning algorithm Laplacian Eigenmaps - LEM, which was chosen due to the mathematical
formulation allows to add information to the mapping process. The main goal is to compute
the relationships among samples of different datasets based on an intra manifold compari-
son to unfold properly the data underlying structure taking into account that we are given
multiple data sets and there is an underlying common manifold among them.
In this sense, we propose to related each input sample xi with C different manifolds
that share a similar underlying structure. Let ΨΨΨ = {Xc}Cc=1 an input manifold set, where
Xc ∈ Rnc×pc . Our goal is to find a mapping from ΨΨΨ to a low-dimensional space Y ∈ Rn×m
(with m� pc, and n =∑C
c=1 nc), which reveals both the intra manifold structure (relation-
ships within manifold), and the inter manifold structure (relationships among manifolds).
Consequently, a weight matrix A that takes into account both structures can be computed
as
A =
W1 M12 · · · M1c · · · M1C
M21 W2 · · · M2c · · · M2C
......
. . ....
. . ....
Mc1 Mc2 · · · Wc · · · McC
......
. . ....
. . ....
MC1 MC2 · · · MCc · · · WC
, (4-1)
where each Wc ∈ Rnc×nc is the traditional LEM intra manifold weight matrix for each Xc
[33]. Furthermore, each Mcb ∈ Rnc×nb (b = 1, . . . , C) block is a soft correspondence matrix
Page 16
40 4 Multiple Manifold Learning for Dimensionality Reduction
between Xc and Xb.
In a previous work [33], is proposed a methodology called Learning a Joint Manifold
Representation (LJM) to unfold the data underlying structure from multiple manifolds. LJM
calculates the matrix A (equation (4-1)), computing the intra manifold structure matrices
Wc as in traditional LEM, and the inter manifold structure matrices Mcb by solving a
permutation matrix P, which allows to find a maximum weight matching by permuting the
rows of Ucb ∈ Rnc×nb , U cbqr = κ (xq,xr), xq ∈ Xc, and xr ∈ Xb (q = 1, . . . , nc; r = 1, . . . , nb ).
Nonetheless, LJM is sensitive to feature variability between samples of different manifolds,
due to Ucb is inferred in the high-dimensional space. Moreover, LJM is limited to analyze
input matrices Xc which belong to the same input dimension (p1 = p2 = · · · = pc = · · · =
pC), as can be seen in the calculation of each Ucb.
In this work, we identify the correspondence among data points from different manifolds
without making a high-dimensional sample comparison. In other words, the similarities
among observations of different manifolds are not directly calculated in each pair Xc and
Xb. Therefore, we compute each soft correspondence matrix Mcb in (4-1) as
M cbqr =
⟨wcq,w
br
⟩∥∥wcq
∥∥ ‖wbr‖, (4-2)
where wcq ∈ R1×nc and wb
r ∈ R1×nb are row vectors of Wc and Wb, respectively. It is
important to note that equation (4-2) is not well defined when nc 6= nb, thereby, we use
an interpolation method based on cubic splines for oversampling the lowest size vector, to
properly compute the inner product between wcq and wb
r. Our approach for Multiple Manifold
Learning (MML) aims to calculate the relationships among samples of different manifolds,
comparing the intra manifold similarities contained in each Wc (equation (4-2)). Finally,
given the weight matrix A, we minimize the following objective function∑ij
(yi − yj)2Aij. (4-3)
Solving equation (4-3) as in traditional LEM algorithm allows us to find a low-dimensional
space Y for data lying on multiple manifolds.
4.2. Experiments and Results
We tested the conventional LEM algorithm, the state of the art methodology named LJM,
and our proposed methodology called MML on two real-world databases, in order to find a
2-D low-dimensional representation (m = 2) for data lying on multiple manifolds.
The first database, the Columbia Object Image Library (COIL-100) [36], contains 72
RGB-color images, for each one of the 100 objects, in PNG format, which were taken at pose
intervals of 5 degrees while the object is rotated 360 degrees. In this work, the following
Page 17
4.2 Experiments and Results 41
objects are used: Car, Frog and Duck. The image size is 128× 128, which are transformed
to gray scale. In Figure 3-4 are shown some examples.
The second database is the CMU motion of body (Mobo)[37], which holds 25 individuals
walking a treadmill in the CMU 3-D room. The subjects perform four different walk patterns:
slow walk, fast walk, incline walk and walking with a ball. All subjects are captured using
six high resolution color cameras distributed evenly around the treadmill. Each sequence
is 11 seconds long, recorded at 30 frames per second. For concrete testing, we used the
silhouette sequences of one gait cycle for slow walk of three persons(subject 02, 06 and 13),
which are captured from a side view (camera vr03 7). The images are resized to 80 × 61.
Some examples of the silhouette images are shown in Figure 4.2.
(a) COIL-100
(b) COIL-100
Figure 4-1.: Databases examples
Three types of experiments are performed in order to validate the proposed methodology
and show its advantages against other frameworks. Firstly, we use the selected objects of
COIL-100 with a same amount of observations per set (n1 = n2 = n3 = 72), and equal input
dimensions (p1 = p2 = p3 = 16384). In this case, the number of nearest neighbors is fixed as
k = 3. In Figure 4-2 are shown the results for this experiment.
For the second experiment, we use the CMU Mobo database, which leads input samples
per manifold of different sizes: n1 = 36, n2 = 40, n3 = 38 and p1 = p2 = p3 = 4880 according
to each subject. The number of neighbors is set to k = 2. In order to test the algorithms on
a dataset that contains multiple manifolds with high-variability in the input sample sizes,
we use the COIL-100 but performing an uniform sampling of the observations. Therefore,
input spaces with n1 = 72, n2 = 36, n3 = 18 and p1 = p2 = p3 = 16384 were obtained. Here,
the number of nearest neighbors are fixed as k1 = 4, k2 = 2, k3 = 1, taking in to account
Page 18
42 4 Multiple Manifold Learning for Dimensionality Reduction
that is not suitable to set the same value of k because the sample size is so different for each
manifold. The results for the second experiment are presented in Figure 4-3 and 4-4.
Finally, the third experiment aims to validate the proposed methodology for analyzing
datasets with different amount of observations and different input dimensions (or image
resolution). For this purpose, we employ the COIL-100 performing an uniform sampling
of the observations, and resizing the images. Thence, the obtained input spaces have the
following characteristics: n1 = 72, n2 = 36, n3 = 18 and p1 = 16384, p2 = 8100, p3 = 2784.
This experiment is only carried out for the proposed approach, due to the other techniques
are not able to work with different input dimensions (Figure 4-5).
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
(a) Laplacian Eigenmaps
-0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03-0.06
-0.04
-0.02
0
0.02
0.04
(b) Joint Manifold
-0.03 -0.02 -0.01 0 0.01 0.02 0.03
-0.03
-0.02
-0.01
0
0.01
0.02
0.03
(c) Proposed approach - MML (d) Matrix A
Figure 4-2.: Three COIL-100 objects, equal amount of observations
Page 19
4.3 Discussion 43
-0.01 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
(a) Laplacian Eigenmaps
-0.08 -0.06 -0.04 -0.02 0 0.02 0.04 0.06
-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
(b) Joint Manifold
-0.05 0 0.05
-0.04
-0.02
0
0.02
0.04
0.06
0.08
(c) Proposed approach - MML (d) Matrix A
Figure 4-3.: Three COIL-100 objects, different amount of observations
4.3. Discussion
According to the results presented in Figures 4.2(a), 4.3(a) and 4.4(a), the traditional LEM is
not able to find the correspondence among different datasets which are related to a common
underlying data structure. For all the provided experiments, LEM performs a clustering
of points for each manifold and it does not preserve the local or global properties of the
high-dimensional data. This algorithm can not find a low-dimensional representation that
unfolds the underlying data structure from multiple manifolds, due to the weight matrix W
in LEM is computed only considering pixel intensity similarities among frames, which in this
case produce the above mentioned clusters. The performance is the same for the different
experimental conditions and both databases.
Again, taking into the account the attained results with the LJM technique (Figures 4.2(b),
Page 20
44 4 Multiple Manifold Learning for Dimensionality Reduction
0 0.02 0.04 0.06 0.08 0.1 0.12
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
(a) Laplacian Eigenmaps
-0.06 -0.04 -0.02 0 0.02 0.04 0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
(b) Joint Manifold
-0.06 -0.04 -0.02 0 0.02 0.04
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
(c) Proposed approach - MML (d) Matrix A
Figure 4-4.: Three Gait subjects, different amount of observations
4.3(b), and 4.4(b)), it can be seen how it attempts to find a correspondence among datasets,
nevertheless the intrinsic geometry data structure of the phenomenon (object motion) is lost.
More precisely, for the COIL-100 database, the dynamic of the rotation is not reflected in
the embedded space. Similar results are obtained for gait analysis in the Mobo database,
although LJM tries to reveal the elliptical motion shape, it is not able to conserve a soft
correspondence among samples. Note that the application of LJM technique is limited to
analyze frames of video sharing a similar geometry, due to Ucb is inferred in the high-
dimensional space (pixels frame comparison). Overall, the LJM method can not properly
learn the relationships among objects/subjects performing the same activity, it just develops
well when the analyzed manifolds are similar in appearance, which limits the applicability
of the technique.
Page 21
4.3 Discussion 45
(a) Multi Manifold Embedding
(b) Matrix A
Figure 4-5.: Different amount of observations and input dimensions (image resolution)