=Paper=
{{Paper
|id=Vol-1175/CLEF2009wn-ImageCLEF-GlotinEt2009b
|storemode=property
|title=Fast LSIS Profile Entropy Features for Robot Visual Self-Localization
|pdfUrl=https://ceur-ws.org/Vol-1175/CLEF2009wn-ImageCLEF-GlotinEt2009b.pdf
|volume=Vol-1175
|dblpUrl=https://dblp.org/rec/conf/clef/GlotinZD09
}}
==Fast LSIS Profile Entropy Features for Robot Visual Self-Localization==
https://ceur-ws.org/Vol-1175/CLEF2009wn-ImageCLEF-GlotinEt2009b.pdf
Fast LSIS Profile Entropy Features for Robot
Visual Self-Localization
Herve GLOTIN1,2 , Zhong-Qiu ZHAO1,3 , Emilie DUMONT1
1
Lab. Sciences de l’Information et des Systmes LSIS, UMR USTV CNRS 6168 La Garde-France
2
University of Sud Toulon-Var, France
3
School of Computer & Information, Hefei University of Technology, China
glotin@univ-tln.fr; zhongqiuzhao@gmail.com
Abstract
In the Robot Vision task, the participants are asked to answer where is the robot
using its vision. The robot may be in 5 rooms (BO-One-person office, CR-Corridor,
EO-Two-persons office, KT-Kitchen, PA-Printer Area). In order to train our models
we structured the views of each room into several sub-classes: BO-inside, exit, enter;
CR-enter, exit, leftstairs, nostairs, rightstairs; EO-enter, exit, inside; KT-enter, exit,
cooking hearth, table, television; PA-cabinet, enter, exit. Then an SVM was con-
structed for each of these 19 sub-classes. After that, we combined the results of SVMs
by maximizing to get the final decision. We run our classification models on the new
Profile Entropy Features (PEF) that combines RGB color and texture, yielding to one
hundred of dimension, and we compare them to generic Descriptor of Fourier (DF).
We also made a fusion of the models on these 2 different features. So we got 3 runs.
In our experiments, for each decision, we used only the current image, but we do not
exploit continuity of the sequences. For this case, a total of 7 teams submitted runs.
The official evaluation give for the SVM(PEF) run a score of 544, and for SVM(DF)
run a score of -32, while their fusion a score of 509.5. Thus our result possesses the
5th rank over the seven. The experiments show that our SVM model works well with
little training cost, and PEF feature works much better than DF feature. It could
be concluded that PEF is quite efficient: it is very fast to be computed, with around
10 images computed per second on usual pentium, and less of 2 hours of training
(compared to 60 hours for the best systems), but still give a competitive results.
Categories and Subject Descriptors
H.3 [Information Storage and Retrieval]: H.3.1 Content Analysis and Indexing; H.3.3 Infor-
mation Search and Retrieval; H.3.4 Systems and Software; H.3.7 Digital Libraries; H.2.3 [Database
Managment]: Visual Systems—CBIR
Keywords
Efficient visual features ; Fast Content Based Information Retrieval ; Profile Entropy Features ;
LS-SVM ; Robot Vision ; Adaptive Localisation
1 Introduction
The task addresses the problem of topological localization of a mobile robot using visual informa-
tion. Specifically, participants will be asked to determine the topological location of a robot based
on images acquired with a perspective camera mounted on a robot platform. The details of this
task and dataset are shown in [10].
We manually classified the views of each room into several classes: BO-inside, exit, enter; CR-
enter, exit, leftstairs, nostairs, rightstairs; EO-enter, exit, inside; KT-enter, exit, cooking hearth,
table, television; PA-cabinet, enter, exit. Then an SVM was constructed for each small class,
to distinguish this small-calss from all the others and recognize if the robot is currently in this
small-class or not. So we constructed 19 SVMs in total(one per class). After that, we combined
the results of SVMs by maximizing to get the final decision. We run our classification models on
PEF(Profile Entropy Feature) and DF(Descriptor of Fourier) features and compared them, and
we also made a fusion of the models on these 2 different features. So we submitted 3 runs. In our
experiments, for each decision, we used only the current image, but not exploit continuity of the
sequences.
2 The Profile Entropy Features
We propose in [1, 2] a new feature equal to the pixel ’profile’ entropy. A pixel profile can be a
simple arithmetic mean in horizontal (or vertical) direction. The advantage of such feature is to
combine raw shape and texture representations in a low cpu cost feature.
Let I be an image (or a part of) of L(I) rows, and C(I) columns. The PEF are computed on
these normalized RGB channels : l = (R + G + B)/3, r = R/l, and g = G/l. We consider the
profiles of the orthogonal projections of the pixels to the horizontal X axis, noted Π op X , and to the
vertical Y axis (Πop
Y ), where op
is a projection operator. This one is either the arithmetic mean of
the pixels (noted ΠAri
. ), or their harmonic mean (noted Π Harm
. ), as illustrated in Fig.1 and Fig.2.
Thus the length of a given profile is either S = C(I) or S = L(I).
Then, for√each profile, we estimate its probability distribution function (pdf ˆ ) on N bins (where
N = round( S) as proposed in [3]).
For each channel, and each operator op , we compute :
Φop ˆ op
X (I) = pdf (ΠX (I)). Considering that the sources are ergodic, we set P EF X component to the
normalised entropy of this distribution :
P EFX (I) = H(Φop X (I))/log(N ),
where N the number of bins of the considered distribution, and H the usual entropy function. We
compute similar PEF on Y axis :
P EFY (I) = H(Φop Y (I))/log(N ).
We set a third PEF component to the entropy of the direct distribution of all the pixels in I,
ˆ (I) :
pdf
ˆ
P EFB (I) = H(pdf(I))/log(N ),
p
where N = round( L(I) ∗ C(I)) bins.
The whole PEF features are the concatenation of P EFX , P EFY and P EFB , with the usual
mean and standard deviation of each channel of I.
The PEF are computed on three horizontal (noted ’=’) or vertical (’kk’ ) equal segmented
subimages, and on the whole image. For exemple, for a given operator, we have the whole image
plus the three ’kk’ subimages, and for each of the 3 channels we have P EFX ,Y ,B , plus their mean
and variance, thus we have 4 ∗ 3 ∗ (3 + 2) = 60 dimensions. We note ’#’ the concatenation of ’=’
and ’kk’ PEF, without duplication.
3 Fast classification using Least Squares Support Vector
Machines
In order to design fast image retrieval systems, we use the Least Squares Support Vector Machine
(LS-SVM). The SVM [5] first maps the data into a higher dimensional input space by some kernel
Figure 1: Horizontal X (Bottom) and, vertical Y (Top Right) profiles using arithmetic (-.-) and
harmonic (-) projections of the luminance of an image of a tree. It shows clearly the difference
between the two projections for this structured pattern.
functions, and then learns a separating hyperspace to maximize the margin. Currently, because
of its good generalization capability, this technique has been widely applied in many areas such
as face detection, image retrieval, and so on [6, 7]. The SVM is typically based on an ε-insensitive
cost function, meaning that approximation errors smaller than ε will not increase the cost function
value. This results in a quadratic convex optimization problem. So instead of using an ε-insensitive
cost function, a quadratic cost function can be used. The least squares support vector machines
(LS-SVM) [8] are reformulations to the standard SVMs which lead to solving linear KKT systems
instead, which is quite computationally attractive. Thus, in all our experiments, we will use the
LS-SVMlab1.5 (http://www.esat.kuleuven.ac.be/sista/lssvmlab/).
In our experiments, the RBF kernel
K(x1 − x2 ) = exp(−|x1 − x2 |2 /σ 2 )
is selected as the kernel function of our LS-SVM. So there is a corresponding parameter, σ , to be
tuned. A large value of σ 2 indicates a stronger smoothing. Moreover, there is another parameter,
γ, needing tuning to find the tradeoff between to stress minimizing of the complexity of the model
and to stress good fitting of the training data points.
We set these two parameters as
σ 2 = [4 25 100 400 600 800 1000 2000]
and
γ = [4 8 16 32 64 128 256 512]
respectively. So hundred of SVMs were constructed for each SVM model, and then we selected
the best SVM using the validation set.
4 Experimental Results
In this task, the participants are asked to answer ’where are the robots’. The robots may be
in 5 rooms (BO-One-person office, CR-Corridor, EO-Two-persons office, KT-Kitchen, PA-Printer
Area). We manually classified the views of each room into several classes: BO-inside, exit, en-
Figure 2: Similar to Fig.1 but for an image of the concept sky : arithmetic and harmonic profiles
are similar.
Table 1: The submitted runs of ImageCLEF2009 Robot Vision Task
Rank Run Tag Method Score Training time
1 MIRG UG matched points; illumination filter 890.5 unknown
2 Idiap CRFH+SIFT+PACT combined using G-DAS 793.0 unknown
3 UAIC Full search using all frames 787.0 60h
4 CVIUI2R LAB+histograms+Probabilistic SVM 784.0 unknown
5 LSIS 1 PEF+SVM 544.0 1h
6 SIMD Hough+Canny+SIFT 511.0 60h
7 LSIS 3 Fusion of (PEF+SVM)(DF+SVM) 509.5 2h
8 MRIM 5x5 patches, Color SIFT, Language Model 456.5 unknown
9 LSIS 2 DF+SVM -32.0 1h
ter; CR-enter, exit, leftstairs, nostairs, rightstairs; EO-enter, exit, inside; KT-enter, exit, cook-
ing hearth, table, television; PA-cabinet, enter, exit (Figure 3). Then an SVM was constructed
for each small class, to distinguish this small-calss from all the others and recognize if the robot is
currently in this small-class or not. So we constructed 19 SVMs in total. After that, we combined
the results of SVMs by maximizing to get the final decision. We run our classification models
on PEF(Profile Entropy Feature) and DF [9] features and compared them, and we also made a
fusion of the models on these 2 different features. So we got 3 runs. In our experiments, for each
definition, we used only the current image, but not exploit continuity of the sequences.
The evaluation was performed by the organizer. The following rules are used when calculating
the score for a single test sequence: (1). +1.0 points for each correctly classified image Correct
detection of an unknown room is treated the same way as correct classification. (2). -0.5 points
for each misclassified image (3). 0.0 points for each image that was not classified (the algorithm
refrained from the decision) In case several test sequences are used, the scores are calculated
separately for each test sequence and then summarized.
As the evaluation, we got the SVM+PEF run with the score of 544, and the SVM+DF run
with the score of -32, and the fusion run with the score of 509.5 (Table 4). The best result of us
possesses the 9th rank. The experiments show that our SVM model works well with little training
cost, and PEF feature works much better than DF feature. It could be concluded that PEF is
efficient. Moreover, PEF is fast with around 10 images computed per second on usual pentium.
Figure 3: The framework of our system
5 Conclusion
Our SVM model on PEF works well, with medium performances, needing much less training time
than other systems (around 50 times less). It could be concluded that our system with PEF feature
is efficient for this task. We also noticed that the best result of those who used the continuity
information, which is from SIMD group, attains a much higher score: 916.5. So in the future, we
will combine in our system the continuity information through HMM optimisation, which shall
results in a significant enhancement.
Acknowledgment
This work was supported by French National Agency of Research (ANR-06-MDCA-002) and Re-
search Fund for the Doctoral Program of Higher Education of China (200803591024).
References
[1] Glotin, H.: Information retrieval and robust perception for a scaled multi-structuration, Thesis
for habilitation of research direction, University Sud Toulon-Var, Toulon (2007)
[2] Glotin, H., Zhao, Z.Q., Ayache, S.: Efficient Image Concept Indexing by Harmonic & Arith-
metic Profiles Entropy, 2009 IEEE International Conference on Image Processing, Cairo, Egypt,
November 7-11, 2009 (2009)
[3] Moddemeijer, R.: On estimation of entropy and mutual information of continuous distribu-
tions, Signal Processing, 16(3), 233–246 (1989)
[4] Vapnik, V.: The nature of statistical learning theory. Springer-Verlag, New York (1995)
[5] Vapnik, V.: Statistical learning theory. John Wiley, New York (1998)
[6] Waring, C.A., Liu, X.: Face detection using spectral histograms and SVMs. IEEE Transactions
on Systems, Man, and Cybernetics, Part B, 35(3), 467–476 (2005)
[7] Tong S., Edward, Chang: Support vector machine active learning for image retrieval. In Pro-
ceedings of the ninth ACM international conference on Multimedia Ottawa, Canada, pp. 107–118
(2001)
[8] Suykens, J.A.K., Vandewalle, J.: Least Squares Support Vector Machine Classifiers Neural
Processing Letters, 9, 293–300 (1999)
[9] Smach, F., Lemaitre, C., Gauthier, J.P., Miteran, J., Atri, M.: Generalized Fourier Descriptors
with Applications to Objects Recognition in SVM Context, 30, J. Math Imaging Vis 43–71
(2008)
[10] Caputo B., Pronobis A., Jensfelt, P.: Overview of the CLEF 2009 robot vision track, CLEF
working notes 2009, Corfu, Greece (2009)