ICPRAM 2012 Abstracts


Area 1 - Theory and Methods

Full Papers
Paper Nr: 9
Title:

ADAPTATION AND ENHANCEMENT OF EVALUATION MEASURES TO OVERLAPPING GRAPH CLUSTERINGS

Authors:

Tatiana Gossen, Michael Kotzyba and Andreas Nürnberger

Abstract: Quality measures are important to evaluate graph clustering algorithms by providing a means to assess the quality of a derived cluster structure. In this paper, we focus on overlapping graph structures, as many realworld networks have a structure of highly overlapping cohesive groups. We propose three methods to adapt existing crisp quality measures such that they can handle graph overlaps correctly, but also ensure that their properties for the evaluation of crisp graph clusterings are preserved when assessing a crisp cluster structure. We demonstrate our methods on such measures as Density, Newman’s modularity and Conductance. We also propose an enhancement of an existing modularity measure for networks with overlapping structure. The newly proposed measures are analysed using experiments on artificial graphs that possess overlapping structure. For this evaluation, we apply a graph generation model that creates clustered graphs with overlaps that are similar to real-world networks i.e. their node degree and cluster size distribution follow a power law.

Paper Nr: 23
Title:

ON ORDER EQUIVALENCES BETWEEN DISTANCE AND SIMILARITY MEASURES ON SEQUENCES AND TREES

Authors:

Martin Emms and Hector-Hugo Franco-Penya

Abstract: Both ’distance’ and ’similarity’ measures have been proposed for the comparison of sequences and for the comparison of trees, based on scoring mappings, and the paper concerns the equivalence or otherwise of these. These measures are usually parameterised by an atomic ’cost’ table, defining label-dependent values for swaps, deletions and insertions. We look at the question of whether orderings induced by a ’distance’ measure, with some cost-table, can be dualized by a ’similarity’ measure, with some other cost-table, and vice-versa. Three kinds of orderings are considered: alignment-orderings, for fixed source S and target T, neighbour-orderings, where for a fixed S, varying candidate neighbours Ti are ranked, and pair-orderings, where for varying Si, and varying Tj , the pairings hSi,Tji are ranked. We show that (1) alignment-orderings by distance can be dualized by similarity, and vice-versa; (2) neigbour-ordering and pair-ordering by distance can be dualized by similarity; (3) neighbour-ordering and pair-ordering by similarity can sometimes not be dualized by distance. A consequence if this is that there are categorisation and hierarchical clustering outcomes which can be achieved via similarity but not via distance.

Paper Nr: 40
Title:

SCALABLE CORPUS ANNOTATION BY GRAPH CONSTRUCTION AND LABEL PROPAGATION

Authors:

Thomas Lansdall-Welfare, Ilias Flaounas and Nello Cristianini

Abstract: The efficient annotation of documents in vast corpora calls for scalable methods of text classification. Representing the documents in the form of graph vertices, rather than in the form of vectors in a bag of words space, allows for the necessary information to be pre-computed and stored. It also fundamentally changes the problem definition, from a content-based to a relation-based classification problem. Efficiently creating a graph where nearby documents are likely to have the same annotation is the central task of this paper. We compare the effectiveness of various approaches to graph construction by building graphs of 800,000 vertices based on the Reuters corpus, showing that relation-based classification is competitive with Support VectorMachines, which can be considered as state of the art. We further show that the combination of our relation-based approach and Support Vector Machines leads to an improvement over the methods individually.

Paper Nr: 73
Title:

NONLINEAR FEATURE CONSTRUCTION WITH EVOLVED NEURAL NETWORKS FOR CLASSIFICATION PROBLEMS

Authors:

Tobias Berka and Helmut A. Mayer

Abstract: Predicting the class membership of a set of patterns represented by points in a multi-dimensional space critically depends on their specific distribution. To improve the classification performance, pattern vectors may be transformed. There is a range of linear methods for feature construction, but these are often limited in their performance. Nonlinear methods are a more recent development in this field, but these pose difficult optimization problems. Evolutionary approaches have been used to optimize both linear and nonlinear functions for feature construction. For nonlinear feature construction, a particular problem is how to encode the function in order to limit the huge search space while preserving enough flexibility to evolve effective solutions. In this paper, we present a new method for generating a nonlinear function for feature construction using multi-layer perceptrons whose weights are shaped by evolution. By pre-defining the architecture of the neural network we can directly influence the computational capacity of the function and the number of features to be constructed. We evaluate the suggested neural feature construction on four commonly used data sets and report an improvement in classification accuracy ranging from 4 to 13 percentage points over the performance on the original pattern set.

Paper Nr: 76
Title:

SPARSE QUASI-NEWTON OPTIMIZATION FOR SEMI-SUPERVISED SUPPORT VECTOR MACHINES

Authors:

Fabian Gieseke, Antti Airola, Tapio Pahikkala and Oliver Kramer

Abstract: In real-world scenarios, labeled data is often rare while unlabeled data can be obtained in huge quantities. A current research direction in machine learning is the concept of semi-supervised support vector machines. This type of binary classification approach aims at taking the additional information provided by unlabeled patterns into account to reveal more information about the structure of the data and, hence, to yield models with a better classification performance. However, generating these semi-supervised models requires solving difficult optimization tasks. In this work, we present a simple but effective approach to address the induced optimization task, which is based on a special instance of the quasi-Newton family of optimization schemes. The resulting framework can be implemented easily using black box optimization engines and yields excellent classification and runtime results on both artificial and real-world data sets that are superior (or at least competitive) to the ones obtained by competing state-of-the-art methods.

Paper Nr: 92
Title:

IRIS RECOGNITION IN VISIBLE LIGHT DOMAIN

Authors:

Daniel Riccio and Maria De Marsico

Abstract: Present iris recognition techniques allow very high recognition performances in controlled settings and with cooperating users; this makes iris a real competitor to other biometric traits like fingerprints, with the further advantage of requiring a contactless acquisition. Moreover, most of the existing approaches are designed for Near Infrared or Hyperspectral images, which are less affected by changes in illumination conditions. Current research is focusing on designing new techniques aiming to ensure high accuracy even on images acquired in visible light and in adverse conditions. This paper deals with an approach to iris matching based on the combination of local features: Linear Binary Patterns (LBP) and discriminable textons (BLOBs). Both these technique have been readapted in order to deal with images captured in variable visible light conditions, and affected by noise due to distance/resolution or to scarce user collaboration (blurring, off-axis iris, occlusion by eyelashes and eyelids). The obtained results are quite convincing and strongly motivate the addition of more local features.

Paper Nr: 98
Title:

NONLINEAR MAPPING BY CONSTRAINED CO-CLUSTERING

Authors:

Rodolphe Priam, Mohamed Nadif and Gérard Govaert

Abstract: The latent block model is an efficient alternative to the mixture model for modelling a dataset when the number of rows or columns of the data matrix studied is large. For analyzing and reducing the spaces of a matrix, the methods proposed in the litterature are most of the time with their foundation in a non-parametric or a mixture model approach. We present an embedding of the projection of co-occurrence tables in the Poisson latent block mixture model. Our approach leads to an efficient way to cluster and reduce this kind of data matrices.

Paper Nr: 116
Title:

CONVEX COMBINATIONS OF MAXIMUM MARGIN BAYESIAN NETWORK CLASSIFIERS

Authors:

Sebastian Tschiatschek and Franz Pernkopf

Abstract: Maximum margin Bayesian networks (MMBN) can be trained by solving a convex optimization problem using, for example, interior point (IP) methods (Guo et al., 2005). However, for large datasets this training is computationally expensive (in terms of runtime and memory requirements). Therefore, we propose a less resource intensive batch method to approximately learn a MMBN classifier: we train a set of (weak) MMBN classifiers on subsets of the training data, and then exploit the convexity of the original optimization problem to obtain an approximate solution, i.e., we determine a convex combination of the weak classifiers. In experiments on different datasets we obtain similar results as for optimal MMBN determined on all training samples. However, in terms of computational efficiency (runtime) we are faster and the memory requirements are much lower. Further, the proposed method facilitates parallel implementation.

Paper Nr: 129
Title:

ESTIMATION OF THE COMMON OSCILLATION FOR PHASE LOCKED MATRIX FACTORIZATION

Authors:

Miguel Almeida, Ricardo Vigário and José Bioucas-Dias

Abstract: Phase Locked Matrix Factorization (PLMF) is an algorithm to perform separation of synchronous sources. Such a problem cannot be addressed by orthodox methods such as Independent Component Analysis, because synchronous sources are highly mutually dependent. PLMF separates available data into the mixing matrix and the sources; the sources are then decomposed into amplitude and phase components. Previously, PLMF was applicable only if the oscillatory component, common to all synchronized sources, was known, which is clearly a restrictive assumption. The main goal of this paper is to present a version of PLMF where this assumption is no longer needed – the oscillatory component can be estimated alongside all the other variables, thus making PLMF much more applicable to real-world data. Furthermore, the optimization procedures in the original PLMF are improved. Results on simulated data illustrate that this new approach successfully estimates the oscillatory component, together with the remaining variables, showing that the general problem of separation of synchronous sources can now be tackled.

Paper Nr: 155
Title:

A GENERAL ALGORITHM FOR CALCULATING FORCE HISTOGRAMS USING VECTOR DATA

Authors:

Daniel Recoskie, Tao Xu and Pascal Matsakis

Abstract: The histogram of forces is a generic relative position descriptor with remarkable properties, and it has found many applications, in various domains. So far, however, the applications involve objects in raster form. The fact is that several general algorithms for the computation of force histograms when dealing with such objects have been developed; on the other hand, there is no general algorithm available for objects in vector form, and the algorithms for raster objects cannot be adapted to vector objects. Here, the first general algorithm for calculating force histograms using vector data is presented.

Paper Nr: 163
Title:

ON THE CROSSOVER OPERATOR FOR GA-BASED OPTIMIZERS IN SEQUENTIAL PROJECTION PURSUIT

Authors:

Soledad Espezua, Edwin Villanueva and Carlos D. Maciel

Abstract: Sequential Projection Pursuit (SPP) is a useful tool to uncover structures hidden in high-dimensional data by constructing sequentially the basis of a low-dimensional projection space where the structure is exposed. Genetic algorithms (GAs) are promising finders of optimal basis for SPP, but their performance is determined by the choice of the crossover operator. It is unknown until now which operator is more suitable for SPP. In this paper we compare, over four public datasets, the performance of eight crossover operators: three available in literature (arithmetic, single-point and multi-point) and five new proposed here (two hyperconic, two fitness biased and one extension of arithmetic crossover). The proposed hyperconic operators and the multi-point operator showed the best performance, finding high-fitness projections. However, it was noted that the final selection is dependent on the dataset dimension and the timeframe allowed to get the answer. Some guidelines to select the most appropriate operator for each situation are presented.

Paper Nr: 172
Title:

A DYNAMIC WRAPPER METHOD FOR FEATURE DISCRETIZATION AND SELECTION

Authors:

Artur Ferreira and Mario Figueiredo

Abstract: In many learning problems, an adequate (sometimes discrete) representation of the data is necessary. For instance, for large number of features and small number of instances, learning algorithms may be confronted with the curse of dimensionality, and need to address it in order to be effective. Feature selection and feature discretization techniques have been used to achieve adequate representations of the data, by selecting an adequate subset of features with a convenient representation. In this paper, we propose static and dynamic methods for feature discretization. The static method is unsupervised and the dynamic method uses a wrapper approach with a quantizer and a classifier, and it can be coupled with any static (unsupervised or supervised) discretization procedure. The proposed methods attain efficient representations that are suitable for learning problems. Moreover, using well-known feature selection methods with the features discretized by our methods leads to better accuracy than with the features discretized by other methods or even with the original features.

Paper Nr: 175
Title:

GENERATIVE EMBEDDINGS BASED ON RICIAN MIXTURES - Application to Kernel-based Discriminative Classification of Magnetic Resonance Images

Authors:

Anna C. Carli, Mario A. T. Figueiredo, Manuele Bicego and Vittorio Murino

Abstract: Most approaches to classifier learning for structured objects (such as images or sequences) are based on probabilistic generative models. On the other hand, state-of-the-art classifiers for vectorial data are learned discriminatively. In recent years, these two dual paradigms have been combined via the use of generative embeddings (of which the Fisher kernel is arguably the best known example); these embeddings are mappings from the object space into a fixed dimensional score space, induced by a generative model learned from data, on which a (maybe kernel-based) discriminative approach can then be used. This paper proposes a new semi-parametric approach to build generative embeddings for classification of magnetic resonance images (MRI). Based on the fact that MRI data is well described by Rice distributions, we propose to use Rician mixtures as the underlying generative model, based on which several different generative embeddings are built. These embeddings yield vectorial representations on which kernel-based support vector machines (SVM) can be trained for classification. Concerning the choice of kernel, we adopt the recently proposed nonextensive information theoretic kernels. The methodology proposed was tested on a challenging classification task, which consists in classifying MRI images as belonging to schizophrenic or non-schizophrenic human subjects. The classification is based on a set of regions of interest (ROIs) in each image, with the classifiers corresponding to each ROI being combined via boosting. The experimental results show that the proposed methodology outperforms the previous state-of-the-art methods on the same dataset.

Paper Nr: 177
Title:

SINGLE-FRAME SIGNAL RECOVERY USING A SIMILARITY-PRIOR BASED ON PEARSON TYPE VII MRF

Authors:

Sakinah Ali Pitchay and Ata Kabán

Abstract: We consider the problem of signal reconstruction from noisy observations in a highly under-determined problem setting. Most of previous work does not consider any specific extra information to recover the signal. Here we address this problem by exploiting the similarity between the signal of interest and a consecutive motionless frame. We incorporate this additional information of similarity that is available into a probabilistic image prior based on the Pearson type VII Markov Random Field model. Results on both synthetic and real data of MRI images demonstrate the effectiveness of our method in both compressed setting and classical super-resolution experiments.

Paper Nr: 184
Title:

CLUSTERING COMPLEX MULTIMEDIA OBJECTS USING AN ENSEMBLE APPROACH

Authors:

Ana Isabel Oviedo and Oscar Ortega

Abstract: A complex multimedia object is an information unit composed by multiple media types like text, images, audio and video. Applications related with huge sets of such objects exceed the human capacity to synthesize useful information. The search for similarities and dissimilarities among objects is a task that has been done through clustering analysis, which tries to find groups in unlabeled data sets. Such analysis applied to complex multimedia object sets has a special restriction. The method must analyze the multiple media types present in the objects. This paper proposes a clustering ensemble that jointly assesses several media types present in this kind of objects. The proposed ensemble was applied to cluster webpages, constructing a text and image clustering prototypes. The Hubert’s statistic was used to evaluate the ensemble performance, showing that the proposed method creates clustering structures more similar to the real classification than a joint-feature vector.

Paper Nr: 225
Title:

ON STOCHASTIC TREE DISTANCES AND THEIR TRAINING VIA EXPECTATION-MAXIMISATION

Authors:

Martin Emms

Abstract: Continuing a line of work initiated in (Boyer et al., 2007), the generalisation of stochastic string distance to a stochastic tree distance is considered. We point out some hitherto overlooked necessary modifications to the Zhang/Shasha tree-distance algorithm for all-paths and viterbi variants of this stochastic tree distance. A strategy towards an EM cost-adaptation algorithm for the all-paths distance which was suggested by (Boyer et al., 2007) is shown to overlook necessary ancestry preservation constraints, and an alternative EM costadaptation algorithm for the Viterbi variant is proposed. Experiments are reported on in which a distanceweighted kNN categorisation algorithm is applied to a corpus of categorised tree structures. We show that a 67.7% base-line using standard unit-costs can be improved to 72.5% by the EM cost adaptation algorithm.

Short Papers
Paper Nr: 18
Title:

THE STEPWISE RESPONSE REFINEMENT SCREENER (SRRS) AND ITS APPLICATIONS TO ANALYSIS OF FACTORIAL EXPERIMENTS

Authors:

Frederick Kin Hing Phoa

Abstract: Two-level supersaturated designs are very useful in the screening experiments and the common goal is to identify sparse but dominant active factors with low cost. Recently, a new analysis procedure called the Stepwise Response Refinement Screener (SRRS) method is proposed to screen important effects. This paper extends this method to the two-level nonregular fractional factorial designs. The applications to several reallife examples suggest that the SRRS method is able to retrieve similar results as the existing methods do. Simulation studies show that compared to existing methods in the literature, the SRRS method performs well in terms of the true model identification rate and the average model size.

Paper Nr: 21
Title:

HANDLING IMPRECISE LABELS IN FEATURE SELECTION WITH GRAPH LAPLACIAN

Authors:

Gauthier Doquire and Michel Verleysen

Abstract: Feature selection is a preprocessing step of great importance for a lot of pattern recognition and machine learning applications, including classification. Even if feature selection has been extensively studied for classical problems, very little work has been done to take into account a possible imprecision or uncertainty in the assignment of the class labels. However, such a situation can be encountered frequently in practice, especially when the labels are given by a human expert having some doubts on the exact class value. In this paper, the problem where each possible class for a given sample is associated with a probability is considered. A feature selection criterion based on the theory of graph Laplacian is proposed and its interest is experimentally demonstrated when compared with basic approaches to handle such imprecise labels.

Paper Nr: 22
Title:

SEMI-LOCAL FEATURES FOR THE CLASSIFICATION OF SEGMENTED OBJECTS

Authors:

Robert Sorschag

Abstract: Image features are usually extracted globally from whole images or locally from regions-of-interest. We propose different approaches to extract semi-local features from segmented objects in the context of object detection. The focus lies on the transformation of arbitrarily shaped object segments to image regions that are suitable for the extraction of features like SIFT, Gabor wavelets, and MPEG-7 color features. In this region transformation step, decisions arise about the used region boundary size and about modifications of the object and its background. Amongst others, we compare uniformly colored, blurred and randomly sampled backgrounds versus simple bounding boxes without object-background modifications. An extensive evaluation on the Pascal VOC 2010 segmentation dataset indicates that semi-local features are suitable for this task and that a significant difference exists between different feature extraction methods.

Paper Nr: 35
Title:

A COMPARISON OF MULTIVARIATE MUTUAL INFORMATION ESTIMATORS FOR FEATURE SELECTION

Authors:

Gauthier Doquire and Michel Verleysen

Abstract: Mutual Information estimation is an important task for many data mining and machine learning applications. In particular, many feature selection algorithms make use of the mutual information criterion and could thus benefit greatly from a reliable way to estimate this criterion. More precisely, the multivariate mutual information (computed between multivariate random variables) can naturally be combined with very popular search procedure such as the greedy forward to build a subset of the most relevant features. Estimating the mutual information (especially through density functions estimations) between high-dimensional variables is however a hard task in practice, due to the limited number of available data points for real-world problems. This paper compares different popular mutual information estimators and shows how a nearest neighbors-based estimator largely outperforms its competitors when used with high-dimensional data.

Paper Nr: 55
Title:

GRAPH RECOGNITION BY SERIATION AND FREQUENT SUBSTRUCTURES MINING

Authors:

Lorenzo Livi, Guido Del Vescovo and Antonello Rizzi

Abstract: Many interesting applications of Pattern Recognition techniques can take advantage in dealing with labeled graphs as input patterns. To this aim the most important issue is the definition of a dissimilarity measure between graphs. In this paper we propose a representation technique able to characterize the input graphs as real valued feature vectors, allowing the use of standard classification systems. This procedure consists in two distinct stages. In the first step a labeled graph is transformed into a sequence of its vertices, ordered according to a given criterion. In a second step this sequence is mapped into a real valued vector. To perform the latter stage, we propose a novel Granular Computing procedure searching for frequent substructures, called GRADIS. This algorithm is in charge of the inexact substructures identification and of the embedding of the sequenced graphs using the symbolic histogram approach. Tests have been performed by synthetically generating a set of graph classification problem instances with the aim to measure system performances when dealing with different types of graphs, as well when increasing problem hardness.

Paper Nr: 90
Title:

LANE DETECTION IN PEDESTRIAN MOTION AND ENTROPY-BASED ORDER INDEX

Authors:

Olivier Chabiron, Jérôme Fehrenbach, Pierre Degond, Mehdi Moussaïd, Julien Pettré and Samuel Lemercier

Abstract: This paper proposes a distance measurement between pedestrian trajectories. This distance is used in a clustering method aiming to detect lanes of pedestrians in experimental data. The main ingredient is to take full advantage of the time sequence available. A study of the sensitivity of the clustering to the parameters shows it is possible to choose a stable set of parameters. We also define an order index based on the concept of entropy. The potential of this index is illustrated in the case of pedestrian lane detection.

Paper Nr: 91
Title:

LINEAR PROJECTION METHODS - An Experimental Study for Regression Problems

Authors:

Carlos Pardo-Aguilar, José F. Diez-Pastor, Nicolás García-Pedrajas, Juan J. Rodríguez and César García-Osorio

Abstract: Two contexts may be considered, in which it is of interest to reduce the dimension of a data set. One of these arises when the intention is to mitigate the curse of dimensionality, when the data set will be used for training a data mining algorithm with a heavy computational load. The other is when one wishes to identify the data set attributes that have a stronger relation with either the class, if dealing with a classification problem, or the value to be predicted, if dealing with a regression problem. Recently, various linear regression projection models have been proposed that attempt to conserve those directions that show the highest correlation with the value to be predicted: Localized Slices Inverse Regression, Weighted Principal Component Analysis and Linear Discriminant Analysis for regression. However, the papers that have presented these methods use only a small number of data sets to validate their smooth functioning. In this research, a more exhaustive study is conducted using 30 data sets. Moreover, by applying the ideas behind these methods, a further three new methods are also presented and included in the comparative study; one of which is competitive with the methods recently proposed.

Paper Nr: 138
Title:

ON THE VC-DIMENSION OF UNIVARIATE DECISION TREES

Authors:

Olcay Taner Yildiz

Abstract: In this paper, we give and prove lower bounds of the VC-dimension of the univariate decision tree hypothesis class. The VC-dimension of the univariate decision tree depends on the VC-dimension values of its subtrees and the number of inputs. In our previous work (Aslan et al., 2009), we proposed a search algorithm that calculates the VC-dimension of univariate decision trees exhaustively. Using the experimental results of that work, we show that our VC-dimension bounds are tight. To verify that the VC-dimension bounds are useful, we also use them to get VC-generalization bounds for complexity control using SRM in decision trees, i.e., pruning. Our simulation results shows that SRM-pruning using the VC-dimension bounds finds trees that are more accurate as those pruned using cross-validation.

Paper Nr: 159
Title:

A PRELIMINARY STUDY ON SELECTING THE OPTIMAL CUT POINTS IN DISCRETIZATION BY EVOLUTIONARY ALGORITHMS

Authors:

Salvador García, Victoria López, Julián Luengo, Cristóbal J. Carmona and Francisco Herrera

Abstract: The Discretization, as a data preprocessing technique, has played an important role in many areas such as artificial intelligence, data mining and machine learning. In this paper, we propose the use of evolutionary algorithms to select a subset of cut points that defines the best possible discretization scheme of a data set. First, we identify the boundary points for each input attribute and then we establish the individual representation as the joining of all of them, forming bit-strings based chromosomes. In addition, we consider an inconsistency based fitness function for measuring the quality of the chromosomes during the evolutionary cycle. The CHC model is adopted as evolutionary approach, showing that it can bring higher accuracy to the discretization process. The proposal has been compared with other state-of-the-art and recent discretizers on 20 real data sets and the experiments show that our proposed algorithm generates competitive discretization schemes in terms of accuracy, for both C4.5 and Naive Bayes classifiers, but using a lower number of cut points.

Paper Nr: 165
Title:

OPTIMIZED ALGORITHM FOR LEARNING BAYESIAN NETWORK SUPER-STRUCTURES

Authors:

Edwin Villanueva and Carlos Dias Maciel

Abstract: Estimating super-structures (SS) as structural constraints for learning Bayesian networks (BN) is an important step of scaling up these models to high-dimensional problems. However, the literature has shown a lack of algorithms with an appropriate accuracy for such purpose. The recent Hybrid Parents and Children - HPC (De Morais and Aussem, 2010) has shown an interesting accuracy, but its local design and high computational cost discourage its use as SS estimator. We present here the OptHPC, an optimized version of HPC that implements several optimizations to get an efficient global method for learning SS. We demonstrate through several experiments that OptHPC estimates SS with the same accuracy than HPC in about 30% of the statistical tests used by it. Also, OptHPC showed the most favorable balance sensitivity/specificity and computational cost for use as super-structure estimator when compared to several state-of-the-art methods.

Paper Nr: 168
Title:

ASSET: APPROXIMATE STOCHASTIC SUBGRADIENT ESTIMATION TRAINING FOR SUPPORT VECTOR MACHINES

Authors:

Sangkyun Lee and Stephen Wright

Abstract: Subgradient methods for training support vector machines have been quite successful for solving large-scale and online learning problems. However, they have been restricted to linear kernels and strongly convex formulations. This paper describes efficient subgradient approaches without such limitations, making use of randomized low-dimensional approximations to nonlinear kernels, and minimization of a reduced primal formulation using an algorithm based on robust stochastic approximation, which do not require strong convexity.

Paper Nr: 181
Title:

DIVISIVE MONOTHETIC CLUSTERING FOR INTERVAL AND HISTOGRAM-VALUED DATA

Authors:

Paula Brito and Marie Chavent

Abstract: In this paper we propose a divisive top-down clustering method designed for interval and histogram-valued data. The method provides a hierarchy on a set of objects together with a monothetic characterization of each formed cluster. At each step, a cluster is split so as to minimize intra-cluster dispersion, which is measured using a distance suitable for the considered variable types. The criterion is minimized across the bipartitions induced by a set of binary questions. Since interval-valued variables may be considered a special case of histogram-valued variables, the method applies to data described by either kind of variables, or by variables of both types. An example illustrates the proposed approach.

Paper Nr: 182
Title:

EVALUATION OF NEGENTROPY-BASED CLUSTER VALIDATION TECHNIQUES IN PROBLEMS WITH INCREASING DIMENSIONALITY

Authors:

L. F. Lago-Fernández, G. Martínez-Muñoz, A. M. González and M. A. Sánchez-Montañés

Abstract: The aim of a crisp cluster validity index is to quantify the quality of a given data partition. It allows to select the best partition out of a set of potential ones, and to determine the number of clusters. Recently, negentropy based cluster validation has been introduced. This new approach seems to perform better than other state of the art techniques, and its computation is quite simple. However, like many other cluster validation approaches, it presents problems when some partition regions have a small number of points. Different heuristics have been proposed to cope with this problem. In this article we systematically analyze the performance of different negentropy-based validation approaches, including a new heuristic, in clustering problems of increasing dimensionality, and compare them to reference criteria such as AIC and BIC. Our results on synthetic data suggest that the newly proposed negentropy-based validation strategy can outperform AIC and BIC when the ratio of the number of points to the dimension is not high, which is a very common situation in most real applications.

Paper Nr: 186
Title:

PROTOTYPE SELECTION IN IMBALANCED DATA FOR DISSIMILARITY REPRESENTATION - A Preliminary Study

Authors:

Mónica Millán Giraldo, Vicente García and J. Salvador Sánchez

Abstract: In classification problems, the dissimilarity representation has shown to be more robust than using the feature space. In order to build the dissimilarity space, a representation set of r objects is used. Several methods have been proposed for the selection of a suitable representation set that maximizes the classification performance. A recurring and crucial challenge in pattern recognition and machine learning refers to the class imbalance problem, which has been said to hinder the performance of learning algorithms. In this paper, we carry out a preliminary study that pursues to investigate the effects of several prototype selection schemes when data set are imbalanced, and also to foresee the benefits of selecting the representation set when the class imbalance is handled by resampling the data set. Statistical analysis of experimental results using Friedman test demonstrates that the application of resampling significantly improve the performance classification.

Paper Nr: 193
Title:

SIMPLEX DECOMPOSITIONS USING SVD AND PLSA

Authors:

Madhusudana Shashanka and Michael Giering

Abstract: Probabilistic Latent Semantic Analysis (PLSA) is a popular technique to analyze non-negative data where multinomial distributions underlying every data vector are expressed as linear combinations of a set of basis distributions. These learned basis distributions that characterize the dataset lie on the standard simplex and themselves represent corners of a simplex within which all data approximations lie. In this paper, we describe a novel method to extend the PLSA decomposition where the bases are not constrained to lie on the standard simplex and thus are better able to characterize the data. The locations of PLSA basis distributions on the standard simplex depend on how the dataset is aligned with respect to the standard simplex. If the directions of maximum variance of the dataset are orthogonal to the standard simplex, then the PLSA bases will give a poor representation of the dataset. Our approach overcomes this drawback by utilizing Singular Values Decomposition (SVD) to identify the directions of maximum variance, and transforming the dataset to align these directions parallel to the standard simplex before performing PLSA. The learned PLSA features are then transformed back into the data space. The effectiveness of the proposed approach is demonstrated with experiments on synthetic data.

Paper Nr: 221
Title:

TRACKING SOLUTIONS OF TIME VARYING LINEAR INVERSE PROBLEMS

Authors:

Martin Kleinsteuber and Simon Hawe

Abstract: The reconstruction of a signal from only a few measurements, deconvolving, or denoising are only a few interesting signal processing applications that can be formulated as linear inverse problems. Commonly, one overcomes the ill-posedness of such problems by finding solutions which best match some prior assumptions. These are often sparsity assumptions as in the theory of Compressive Sensing. In this paper, we propose a method to track solutions of linear inverse problems. We assume that the corresponding solutions vary smoothly over time. A discretized Newton flow allows to incorporate the time varying information for tracking and predicting the subsequent solution. This prediction requires to solve a linear system of equation, which is in general computationally cheaper than solving a new inverse problem. It may also serve as an additional prior that takes the smooth variation of the solutions into account, or, as an initial guess for the preceding reconstruction. We exemplify our approach with the reconstruction of a compressively sampled synthetic video sequence.

Posters
Paper Nr: 20
Title:

HANDWRITTEN CHARACTER CLASSIFICATION USING THE HOTSPOT FEATURE EXTRACTION TECHNIQUE

Authors:

Olarik Surinta, Lambert Schomaker and Marco Wiering

Abstract: Feature extraction techniques can be important in character recognition, because they can enhance the efficacy of recognition in comparison to featureless or pixel-based approaches. This study aims to investigate the novel feature extraction technique called the hotspot technique in order to use it for representing handwritten characters and digits. In the hotspot technique, the distance values between the closest black pixels and the hotspots in each direction are used as representation for a character. The hotspot technique is applied to three data sets including Thai handwritten characters (65 classes), Bangla numeric (10 classes), and MNIST (10 classes). The hotspot technique consists of two parameters including the number of hotspots and the number of chain code directions. The data sets are then classified by the k-Nearest Neighbors algorithm using the Euclidean distance as function for computing distances between data points. In this study, the classification rates obtained from the hotspot, mark direction, and direction of chain code techniques are compared. The results revealed that the hotspot technique provides the largest average classification rates.

Paper Nr: 31
Title:

ON IMPROVING SEMI-SUPERVISED MARGINBOOST INCREMENTALLY USING STRONG UNLABELED DATA

Authors:

Thanh-Binh Le and Sang-Woon Kim

Abstract: The aim of this paper is to present an incremental learning strategy by which the classification accuracy of the semi-supervised MarginBoost (SSMB) algorithm (d’Alch ´ e Buc, 2002) can be improved. In SSMB, both a limited number of labeled and a multitude of unlabeled data are utilized to learn a classification model. However, it is also well known that the utilization of the unlabeled data is not always helpful for semi-supervised learning algorithms. To address this concern when dealing with SSMB, in this paper we study a means of selecting only “small” helpful portion of samples from the additional available data. More specifically, this is done by performing SSMB after incrementally reinforcing the given labeled training data with a part of strong unlabeled data; we train the classification model in an incremental fashion by employing a small amount of “strong” samples selected from the unlabeled data per iteration. The proposed scheme is evaluated with well-known benchmark databases, including some UCI data sets, in two approaches: dissimilarity-based classification (DBC) (Pekalska and Duin, 2005) as well as conventional feature-based classification. Our experimental results demonstrate that, compared to previous approaches, it achieves better classification accuracy results.

Paper Nr: 53
Title:

INEXACT GRAPH MATCHING THROUGH GRAPH COVERAGE

Authors:

Lorenzo Livi, Guido Del Vescovo and Antonello Rizzi

Abstract: In this paper we propose a novel inexact graph matching procedure called graph coverage, to be used in supervised and unsupervised data driven modeling systems. It relies on tensor product between graphs, since the resulting product graph is known to be able to encode the similarity of the two input graphs. The graph coverage is defined on the basis of the concept of graph weight, computed on the weighted adjacency matrix of the tensor product graph. We report the experimental results concerning two distinct performance evaluations. Since for practical applications the computing time of any inexact graph matching procedure should be feasible, the first tests have been conceived to measure the average computing time when increasing the average size of a random sample of fully-labeled graphs. The second one aims to evaluating the accuracy of the proposed dissimilarity measure when used as the core of a classification system based on the k-NN rule. Overall the graph coverage shows encouraging results as a dissimilarity measure.

Paper Nr: 59
Title:

A COMPREHENSIVE DATASET FOR EVALUATING APPROACHES OF VARIOUS META-LEARNING TASKS

Authors:

Matthias Reif

Abstract: New approaches in pattern recognition are typically evaluated against standard datasets, e.g. from UCI or StatLib. Using the same and publicly available datasets increases the comparability and reproducibility of evaluations. In the field of meta-learning, the actual dataset for evaluation is created based on multiple other datasets. Unfortunately, no comprehensive dataset for meta-learning is currently publicly available. In this paper, we present a novel and publicly available dataset for meta-learning based on 83 datasets, six classification algorithms, and 49 meta-features. Different target variables like accuracy and training time of the classifiers as well as parameter dependent measures are included as ground-truth information. Therefore, the meta-dataset can be used for various meta-learning tasks, e.g. predicting the accuracy and training time of classifiers or predicting the optimal parameter values. Using the presented meta-dataset, a convincing and comparable evaluation of new meta-learning approaches is possible.

Paper Nr: 63
Title:

TRAFFIC LIGHT RECOGNITION USING CIRCULAR SEPARABILITY FILTER

Authors:

Shodai Horima and Kazunori Onoguchi

Abstract: This paper proposes the camera-based approach to recognize the traffic light for driver assistance. The circular separability filter applied to RGB images extracts the area of the traffic light. The separability has large value in the boundary where the intensity between two areas changes like the step and it doesn't depend on the intensity difference (height of the step). Scanning the circular mask in each RGB image, the separability is calculated. The separability becomes large in an area where a color is homogeneous and a shape is similar to the circle. Therefore, the pixel with large separability is selected as the candidate of the traffic light. Unlike the conventional method which calculates the circularity from the binarized region, the proposed method can identify the traffic light whose outline is indistinct and whose radius is small. At first, the proposed method removes the region where the saturation is low and the brightness is extremely low or high because there is few possibility that the traffic light is included in these regions. Next, the circular mask is scanned in each RGB image captured from the on-vehicle color camera and the separability between the inside circle and the outside ring is calculated. The maximum value of separability calculated in RGB images is selected as the separability of each pixel. Pixels with large separability are detected as the candidate region of the traffic light. Finally, the candidate region around which inactive traffic lamps exist is identified as the traffic light. Experiments recognizing various traffic lights under various weathers and time show the effectiveness of the proposed method.

Paper Nr: 99
Title:

GENERATIVE TOPOGRAPHIC MAPPING AND FACTOR ANALYZERS

Authors:

Rodolphe Priam and Mohamed Nadif

Abstract: By embedding random factors in the Gaussian mixture model (GMM), we propose a new model called faGTM. Our approach is based on a flexible hierarchical prior for a generalization of the generative topographic mapping (GTM) and the mixture of principal components analyzers (MPPCA). The parameters are estimated with expectation-maximization and maximum a posteriori. Empirical experiments show the interest of our proposal.

Paper Nr: 101
Title:

SCALE-INDEPENDENT SPATIO-TEMPORAL STATISTICAL SHAPE REPRESENTATIONS FOR 3D HUMAN ACTION RECOGNITION

Authors:

Marco Körner, Daniel Haase and Joachim Denzler

Abstract: Since depth measuring devices for real-world scenarios became available in the recent past, the use of 3d data now comes more in focus of human action recognition. We propose a scheme for representing human actions in 3d, which is designed to be invariant with respect to the actor’s scale, rotation, and translation. Our approach employs Principal Component Analysis (PCA) as an exemplary technique from the domain of manifold learning. To distinguish actions regarding their execution speed, we include temporal information into our modeling scheme. Experiments performed on the CMU Motion Capture dataset shows promising recognition rates as well as its robustness with respect to noise and incorrect detection of landmarks.

Paper Nr: 123
Title:

CLOTH COVERING AND APPLICATION TO FEATURE EXTRACTION FOR SCRIPT IDENTIFICATION

Authors:

Minwoo Kim and Il-Seok Oh

Abstract: This paper proposes a concept and algorithm of cloth covering. It is a physically-based model which simulates computationally a shape of cloth covering some objects. It has one scale parameter which controls the degree of suppressing fine-scale structures. To show viability of the proposed cloth covering, this paper performed an experiment of script recognition. The results of comparing accuracies of feature extraction using Gaussian and cloth covering showed that the cloth covering is superior to Gaussian.

Paper Nr: 139
Title:

ONLINE SEQUENTIAL LEARNING BASED ON ENHANCED EXTREME LEARNING MACHINE USING LEFT OR RIGHT PSEUDO-INVERSE

Authors:

Weiwei Zong, Yuan Lan and Guang-Bin Huang

Abstract: The latest development (Huang et al., 2011) has shown that better generalization performance can be obtained for extreme learning machine (ELM) by adding a positive value to the diagonal of HTH or HHT , where H is the hidden layer output matrix. This paper further extends this enhanced ELM to online sequential learning mode. An online sequential learning algorithm is proposed for SLFNs and other regularization networks, consisting of two formulas for two kinds of scenarios: when initial training data is of small scale or large scale. Performance of proposed online sequential learning algorithm is demonstrated through six benchmarking data sets for both regression and multi-class classification problems.

Paper Nr: 148
Title:

CLASSIFICATION USING HIGH ORDER DISSIMILARITIES IN NON-EUCLIDEAN SPACES

Authors:

Helena Aidos, Ana Fred and Robert P. W. Duin

Abstract: This paper introduces a novel classification algorithm named MAP-DID. This algorithm combines a maximum a posteriori (MAP) approach using the well-known Gaussian Mixture Model (GMM) method with a term that forces the various Gaussian components within each class to have a common structure. That structure is based on higher-order statistics of the data, through the use of the dissimilarity increments distribution (DID), which contains information regarding the triplets of neighbor points in the data, as opposed to typical pairwise measures, such as the Euclidean distance. We study the performance ofMAP-DID on several synthetic and real datasets and on various non-Euclidean spaces. The results show that MAP-DID outperforms other classifiers and is therefore appropriate for classification of data on such spaces.

Paper Nr: 160
Title:

ON THE SUITABILITY OF NUMERICAL PERFORMANCE MEASURES FOR CLASS IMBALANCE PROBLEMS

Authors:

Vicente García, J. Salvador Sánchez and Ramón A. Mollineda

Abstract: The class imbalance problem has been reported as an important challenge in various fields such as Pattern Recognition, Data Mining and Machine Learning. A less explored research area is related to how to evaluate classifiers on imbalanced data sets. This work analyzes the behaviour of performance measures widely used on imbalanced problems, as well as other metrics recently proposed in the literature. We perform two theoretical analysis based on Pearson correlation and operations for a 2×2 confusion matrix with the aim to show the strengths and weaknesses of those performance metrics in the presence of skewed distributions.

Paper Nr: 220
Title:

DIMENSION REDUCTION BY AN ORTHOGONAL SERIES ESTIMATE OF THE PROBABILISTIC DEPENDENCE MEASURE

Authors:

Wissal Drira, Wissal Neji and Faouzi Ghorbel

Abstract: Here, we intend to introduce a new estimate of the L2 probabilistic dependence measure by Fourier series for 2-dimensional reduction. Its performance is compared to the Fischer Linear Discriminate Analysis (LDA) and the Approximate Chernoff Criterion (ACC) in the mean of classification probability error.

Area 2 - Applications

Full Papers
Paper Nr: 15
Title:

COMPARING LINEAR AND CONVEX RELAXATIONS FOR STEREO AND MOTION

Authors:

Thomas Schoenemann

Abstract: We provide an analysis of several linear programming relaxations for the problems of stereo disparity estimation and motion estimation. The problems are cast as integer linear programs and their relaxations are solved approximately either by block coordinate descent (TRW-S and MPLP) or by smoothing and convex optimization techniques. We include a comparison to graph cuts. Indeed, the best energies are obtained by combining move-based algorithms and relaxation techniques. Our work includes a (slightly novel) tight relaxation for the typical motion regularity term, where we apply a lifting technique and discuss two ways to solve the arising task. We also give techniques to derive reliable lower bounds, an issue that is not obvious for primal relaxation methods, and apply the technique of (Desmet et al., 1992) to a-priori exclude some of the labels. Moreover we investigate techniques to solve linear and convex programming problems via accelerated first order schemes which are becoming more and more widespread in computer vision.

Paper Nr: 17
Title:

MULTIPLE TARGET TRACKING AND IDENTITY LINKING UNDER SPLIT, MERGE AND OCCLUSION OF TARGETS AND OBSERVATIONS

Authors:

José C. Rubio, Joan Serrat and Antonio M. López

Abstract: Multiple object tracking in video sequences is a difficult problem when one has to simultaneously deal with the following realistic conditions: 1) all or most objects share an identical or very similar appearance, 2) objects are imaged at close positions so there is a data association problem which becomes worse when the number of targets is high, 3) the objects to be tracked may lack observations for a short or long interval, for instance because they are not well detected or are being temporally occluded by another non-target object, and 4) their observations may overlap in the images because the objects are very near or the image results from a 2D projection from the 3D scene, giving rise to the merging and subsequently splitting of tracks. This later condition poses the additional problem of maintaining the objects identity when their observations undergo a merge and split. We pose the tracking and identity linking problem as one of inference on a two-layer probabilistic graphical model and show how can it be efficiently solved. Results are assessed on three very different types of video sequences, showing a turbulent flow of particles, bacteria growth and on-coming traffic headlights.

Paper Nr: 30
Title:

LEARNING A VISUAL ATTENTION MODEL FOR ADAPTIVE FAST-FORWARD IN VIDEO SURVEILLANCE

Authors:

Benjamin Höferlin, Hermann Pflüger, Markus Höferlin, Gunther Heidemann and Daniel Weiskopf

Abstract: The focus of visual attention is guided by salient signals in the peripheral field of view (bottom-up) as well as by the relevance feedback of a semantic model (top-down). As a result, humans are able to evaluate new situations very fast, with only a view numbers of fixations. In this paper, we present a learned model for the fast prediction of visual attention in video. We consider bottom-up and memory-less top-down mechanisms of visual attention guidance, and apply the model to video playback-speed adaption. The presented visual attention model is based on rectangle features that are fast to compute and capable of describing the known mechanisms of bottom-up processing, such as motion, contrast, color, symmetry, and others as well as topdown cues, such as face and person detectors. We show that the visual attention model outperforms other recent methods in adaption of video playback-speed.

Paper Nr: 32
Title:

DETECTION OF HASTY STATE BY MEANS OF USING PSYCHOSOMATIC INFORMATION

Authors:

Masahiro Miyaji, Kenji Takagi, Haruki Kawanaka and Koji Oguri

Abstract: We introduced an Internet survey and analyzed driver’s psychosomatic state immediately before traffic incident by using 7 models of traffic accidents. We identified driver’s hasty is one of key factors which may result in traffic accidents. Aiming at the reduction of the number of traffic accidents, we studied to detect hasty state of a driver while driving by way of using the psychosomatic signals of a driver, which were heart rate and useful field of view. Finally we proposed a concept of a function for detecting driver’s states for an intelligent drive support system.

Paper Nr: 39
Title:

WHAT MAKES US CLICK? - Modelling and Predicting the Appeal of News Articles

Authors:

Elena Hensinger, Ilias Flaounas and Nello Cristianini

Abstract: We model readers’ preferences for online news, and use these models to compare different news outlets with each other. The models are based on linear scoring functions, and are inferred by exploiting aggregate behavioural information about readers’ click choices for textual content of six given news outlets over one year of time. We generate one model per outlet, and while not extremely accurate – due to limited information – these models are shown to predict the click choices of readers, as well as to being stable over time. We use those six audience preference models in several ways: to compare how the audiences’ preferences of different outlets relate to each other; to score different news topics with respect to user appeal; to rank a large number of other news outlets with respect to their content appeal to all audiences; and to explain this measure by relating it to other metrics. We discover that UK tabloids and the website of the “People” magazine contain more appealing content for all audiences than broadsheet newspapers, news aggregators and newswires, and that this measure of readers’ preferences correlates with a measure of linguistic subjectivity at the level of outlets.

Paper Nr: 41
Title:

ESTIMATION OF HUMAN ORIENTATION BASED ON SILHOUETTES AND MACHINE LEARNING PRINCIPLES

Authors:

Sébastien Piérard and Marc Van Droogenbroeck

Abstract: Estimating the orientation of the observed person is a crucial task for home entertainment, man-machine interaction, intelligent vehicles, etc. This is possible but complex with a single camera because it only provides one side view. To decrease the sensitivity to color and texture, we use the silhouette to infer the orientation. Under these conditions, we show that the only intrinsic limitation is to confuse the orientation q with the supplementary angle (that is 180º - q), and that the shape descriptor must distinguish between mirrored images. In this paper, the orientation estimation is expressed and solved in the terms of a regression problem and supervised learning. In our experiments, we have tested and compared 18 shape descriptors; the best one achieves a mean error of 5:24º. However, because of the intrinsic limitation mentioned above, the range of orientations is limited to 180º. Our method is easy to implement and outperforms existing techniques.

Paper Nr: 54
Title:

CLASSIFICATION OF 3D URBAN SCENES - A Voxel based Approach

Authors:

Ahmad Kamal Aijazi, Paul Checchin and Laurent Trassoudaine

Abstract: In this paper we present a method to classify urban scenes based on a super-voxel segmentation of sparse 3D data. The 3D point cloud is first segmented into voxels, which are then joined together by using a link-chain method rather than the usual region growing algorithm to create objects. These objects are then classified using geometrical models and local descriptors. In order to evaluate our results a new metric is presented, which combines both segmentation and classification results simultaneously. The effects of voxel size and incorporation of RGB color and intensity on the classification results are also discussed.

Paper Nr: 58
Title:

CONSTRAINT-FREE TOPOLOGICAL MAPPING AND PATH PLANNING BY MAXIMA DETECTION OF THE KERNEL SPATIAL CLEARANCE DENSITY

Authors:

Panagiotis Papadakis, Mario Gianni, Matia Pizzoli and Fiora Pirri

Abstract: Asserting the inherent topology of the environment perceived by a robot is a key prerequisite of high-level decision making. This is achieved through the construction of a concise representation of the environment that endows a robot with the ability to operate in a coarse-to-fine strategy. In this paper, we propose a novel topological segmentation method of generic metric maps operating concurrently as a path-planning algorithm. First, we apply a Gaussian Distance Transform on the map that weighs points belonging to free space according to the proximity of the surrounding free area in a noise resilient mode. We define a region as the set of all the points that locally converge to a common point of maximum space clearance and employ a weighed meanshift gradient ascent onto the kernel space clearance density in order to detect the maxima that characterize the regions. The spatial intra-connectivity of each cluster is ensured by allowing only for linearly unobstructed mean-shifts which in parallel serves as a path-planning algorithm by concatenating the consecutive mean-shift vectors of the convergence paths. Experiments on structured and unstructured environments demonstrate the effectiveness and potential of the proposed approach.

Paper Nr: 68
Title:

COMPUTING THE REEB GRAPH FOR TRIANGLE MESHES WITH ACTIVE CONTOURS

Authors:

Laura Brandolini and Marco Piastra

Abstract: This paper illustrates a novel method to compute the Reeb graph for triangle meshes. The algorithm is based on the definition of discrete, active contours as counterparts of continuous level lines. Active contours are made up of edges and vertices with multiple presence and implicitly maintain a faithful representation of the level lines, even in case of coarse meshes with higher genus. This approach gives a great advantage in the identification of the nodes in the Reeb graph and also improves the overall efficiency of the algorithm in that at each step only the information local to the contours and their immediate neighborhood needs to be processed. The validation of functional integrity for the algorithm has been carried out experimentally, with real-world data, without mesh pre-processing.

Paper Nr: 69
Title:

AN INTEGRATED APPROACH TO CONTEXTUAL FACE DETECTION

Authors:

Santi Seguí, Michal Drozdzal, Petia Radeva and Jordi Vitrià

Abstract: Face detection is, in general, based on content-based detectors. Nevertheless, the face is a non-rigid object with well defined relations with respect to the human body parts. In this paper, we propose to take benefit of the context information in order to improve content-based face detections. We propose a novel framework for integrating multiple content- and context-based detectors in a discriminative way. Moreover, we develop an integrated scoring procedure that measures the ’faceness’ of each hypothesis and is used to discriminate the detection results. Our approach detects a higher rate of faces while minimizing the number of false detections, giving an average increase of more than 10% in average precision when comparing it to state-of-the art face detectors.

Paper Nr: 70
Title:

COST SENSITIVE AND PREPROCESSING FOR CLASSIFICATION WITH IMBALANCED DATA-SETS: SIMILAR BEHAVIOUR AND POTENTIAL HYBRIDIZATIONS

Authors:

Victoria López, Alberto Fernández, María José del Jesus and Francisco Herrera

Abstract: The scenario of classification with imbalanced data-sets has supposed a serious challenge for researchers along the last years. The main handicap is related to the large number of real applications in which one of the classes of the problem has a few number of examples in comparison with the other class, making it harder to be correctly learnt and, what is most important, this minority class is usually the one with the highest interest. In order to address this problem, two main methodologies have been proposed for stressing the significance of the minority class and for achieving a good discrimination for both classes, namely preprocessing of instances and cost-sensitive learning. The former rebalances the instances of both classes by replicating or creating new instances of the minority class (oversampling) or by removing some instances of the majority class (undersampling); whereas the latter assumes higher misclassification costs with samples in the minority class and seek to minimize the high cost errors. Both solutions have shown to be valid for dealing with the class imbalance problem but, to the best of our knowledge, no comparison between both approaches have ever been performed. In this work, we carry out a full exhaustive analysis on this two methodologies, also including a hybrid procedure that tries to combine the best of these models. We will show, by means of a statistical comparative analysis developed with a large collection of more than 60 imbalanced data-sets, that we cannot highlight an unique approach among the rest, and we will discuss as a potential research line the use of hybridizations for achieving better solutions to the imbalanced data-set problem.

Paper Nr: 81
Title:

AUTOMATIC HOVERFLY SPECIES DISCRIMINATION

Authors:

Branko Brkljač, Marko Panić, Dubravko Ćulibrk, Vladimir Crnojević, Jelena Ačanski and Ante Vujić

Abstract: An novel approach to automatic hoverfly species discrimination based on detection and extraction of vein junctions in wing venation patterns of insects is presented in the paper. The dataset used in our experiments consists of high resolution microscopic wing images of several hoverfly species collected over a relatively long period of time at different geographic locations. Junctions are detected using histograms of oriented gradients and local binary patterns features. The features are used to train an SVM classifier to detect junctions in wing images. Once the junctions are identified they are used to extract simple statistics concerning the distances of these points from the centroid. Such simple features can be used to achieve automatic discrimination of four selected hoverfly species, using a Multi Layer Perceptron (MLP) neural network classifier. The proposed approach achieves classification accuracy of environ 71%.

Paper Nr: 85
Title:

NON-PARAMETRIC SEGMENTATION OF REGIME-SWITCHING TIME SERIES WITH OBLIQUE SWITCHING TREES

Authors:

Alexei Bocharov and Bo Thiesson

Abstract: We introduce a non-parametric approach for the segmentation in regime-switching time-series models. The approach is based on spectral clustering of target-regressor tuples and derives a switching regression tree, where regime switches are modeled by oblique splits. Our segmentation method is very parsimonious in the number of splits evaluated during the construction process of the tree–for a candidate node, the method only proposes one oblique split on regressors and a few targeted splits on time. The regime-switching model can therefore be learned efficiently from data. We use the class of ART time series models to serve as illustration, but because of the non-parametric nature of our segmentation approach, it readily generalizes to a wide range of time-series models that go beyond the Gaussian error assumption in ART models. Experimental results on S&P 1500 financial trading data demonstrates dramatically improved predictive accuracy for the exemplifying ART models.

Paper Nr: 86
Title:

LINE HISTOGRAM - A Fast Method for Rotated Rectangular Area Histogramming

Authors:

Zhiqiang Hou, Sion Hannuna, Xianghua Xie and Majid Mirmehdi

Abstract: We propose a novel approach to executing exhaustive histogram search, which incorporates a process to rotate the template used in a computationally efficient manner, e.g. for tracking applications where object rotation will require template rotation. The method is particularly applicable to rectangular templates as any rotation in the region of interest will greatly reduce the potential for a satisfactory match. We present a computational analysis of our proposed method, followed by comparative experimental results.

Paper Nr: 109
Title:

IMPROVING ELECTRIC FRAUD DETECTION USING CLASS IMBALANCE STRATEGIES

Authors:

Matías Di Martino, Federico Decia, Juan Molinelli and Alicia Fernández

Abstract: Improving nontechnical loss detection is a huge challenge for electric companies. The great number of clients and the diversity of the different types of fraud makes this a very complex task. In this paper we present a fraud detection strategy based on class imbalance research. An automatic detection tool combining classification strategies is proposed. Individual classifiers such as One Class SVM, Cost Sensitive SVM (CS-SVM), Optimum Path Forest (OPF) and C4.5 Tree, and combination functions are designed taken special care in the data’s class imbalance nature. Analysis over consumers historical kWh load profile data from Uruguayan Electric Company (UTE) shows that using combination and balancing techniques improves automatic detection performance.

Paper Nr: 115
Title:

EXTRACTION OF BLOOD DROPLET FLIGHT TRAJECTORIES FROM VIDEOS FOR FORENSIC ANALYSIS

Authors:

L. A. Zarrabeitia, D. A. Aruliah and F. Z. Qureshi

Abstract: We present a method for extracting three-dimensional flight trajectories of liquid droplets from video data. A high-speed stereo camera pair records videos of experimental reconstructions of projectile impacts and ensuing droplet scattering. After background removal and segmentation of individual droplets in each video frame, we introduce a model-based matching technique to accumulate image paths for individual droplets. Our motion detection algorithm is designed to deal gracefully with the lack of feature points, with the similarity of droplets in shape, size, and color, and with incomplete droplet paths due to noise, occlusions, etc. The final reconstruction algorithm pairs two-dimensional paths accumulated from each of the two cameras’ videos to reconstruct trajectories in three dimensions. The reconstructed droplet trajectories constitute a starting point for a physically accurate model of blood droplet flight for forensic bloodstain pattern analysis.

Paper Nr: 118
Title:

LOW LATENCY RECOGNITION AND REPRODUCTION OF NATURAL GESTURE TRAJECTORIES

Authors:

Ulf Großekathöfer, Amir Sadeghipour, Thomas Lingner, Peter Meinicke, Thomas Hermann and Stefan Kopp

Abstract: In human-machine interaction scenarios, low latency recognition and reproduction is crucial for successful communication. For reproduction of general gesture classes it is important to realize a representation that is insensitive with respect to the variation of performer specific speed development along gesture trajectories. Here, we present an approach to learning of speed-invariant gesture models that provide fast recognition and convenient reproduction of gesture trajectories. We evaluate our gesture model with a data set comprising 520 examples for 48 gesture classes. The results indicate that the model is able to learn gestures from few observations with high accuracy.

Paper Nr: 125
Title:

USER BEHAVIOR RECOGNITION FOR AN AUTOMATIC PROMPTING SYSTEM - A Structured Approach based on Task Analysis

Authors:

Christian Peters, Thomas Hermann and Sven Wachsmuth

Abstract: In this paper, we describe a structured approach for user behavior recognition in an automatic prompting system that assists users with cognitive disabilities in the task of brushing their teeth. We analyze the brushing task using qualitative data analysis. The results are a hierarchical decomposition of the task and the identification of environmental configurations during subtasks. We develop a hierarchical recognition framework based on the results of task analysis: We extract a set of features from multimodal sensors which are discretized into the environmental configuration in terms of states of objects involved in the brushing task. We classify subtasks using a Bayesian Network (BN) classifier and a Bayesian Filtering approach. We compare three variants of the BN using different observation models (IU, NaiveBayes and Holistic) with a maximum-margin classifier (multi-class SVM). We present recognition results on 18 trials with regular users and found the BN with a NaiveBayes observation model to produce the best recognition rates of 84.5% on avg.

Paper Nr: 128
Title:

EVENT DETECTION USING LOG-LINEAR MODELS FOR CORONARY CONTRAST AGENT INJECTIONS

Authors:

Dierck Matern, Alexandru Paul Condurache and Alfred Mertins

Abstract: In this paper, we discuss a method to detect contrast agent injections during Percutaneous Transluminal Coronary Angioplasty that is performed to treat the coronary arteries disease. During this intervention contrast agent is injected to make the vessels visible under X-rays. We aim to detect the moment the injected contrast agent reaches the coronary vessels. For this purpose, we use an algorithm based on log-linear models that are a generalization of Conditional Random Fields and Maximum Entropy Markov Models. We show that this more generally applicable algorithm performs in this case similar to dedicated methods.

Paper Nr: 134
Title:

OBJECT RECOGNITION IN PROBABILISTIC 3-D VOLUMETRIC SCENES

Authors:

Maria I. Restrepo, Brandon A. Mayer and Joseph L. Mundy

Abstract: A new representation of 3-d object appearance from video sequences has been developed over the past several years (Pollard and Mundy, 2007; Pollard, 2008; Crispell, 2010), which combines the ideas of background modeling and volumetric multi-view reconstruction. In this representation, Gaussian mixture models for intensity or color are stored in volumetric units. This 3-d probabilistic volume model, PVM, is learned from a video sequence by an on-line Bayesian updating algorithm. To date, the PVM representation has been applied to video image registration (Crispell et al., 2008), change detection (Pollard and Mundy, 2007) and classification of changes as vehicles in 2-d only (Mundy and Ozcanli, 2009; O¨ zcanli and Mundy, 2010). In this paper, the PVM is used to develop novel viewpoint-independent features of object appearance directly in 3-d. The resulting description is then used in a bag-of-features classification algorithm to recognize buildings, houses, parked cars, parked aircraft and parking lots in aerial scenes collected over Providence, Rhode Island, USA. Two approaches to feature description are described and compared: 1) features derived from a PCA analysis of model neighborhoods; and 2) features derived from the coefficients of a 3-d Taylor series expansion within each neighborhood. It is shown that both feature types explain the data with similar accuracy. Finally, the effectiveness of both feature types for recognition is compared for the different categories. Encouraging experimental results demonstrate the descriptive power of the PVM representation for object recognition tasks, promising successful extension to more complex recognition systems.

Paper Nr: 142
Title:

BROADCAST NEWS PHONEME RECOGNITION BY SPARSE CODING

Authors:

Joseph Razik, Sébastien Paris and Hervé Glotin

Abstract: We present in this paper a novel approach for the phoneme recognition task that we want to extend to an automatic speech recognition system (ASR). Usual ASR systems are based on a GMM-HMM combination that represents a fully generative approach. Current discriminative methods are not tractable in large scale data set case, especially with non-linear kernel. In our system, we introduce a new scheme using jointly sparse coding and an approximation additive kernel for fast SVM training for phoneme recognition. Thus, on a broadcast news corpus, our system outperforms the use of GMMs by around 2.5% and is computationally linear to the number of samples.

Paper Nr: 145
Title:

FAST TEMPLATE MATCHING OF REPETITIVE OBJECTS IN STEREOSCOPY

Authors:

Youval Nehmadi, Orly Kalantyrsky and Hugo Guterman

Abstract: One of the challenges of stereovision is to process images with repetitive objects. In order to calculate the distance to an object, matching of the corresponding points between two images must be done. When repetitive objects exist, matching is not straightforward. Many known stereo methods rely on a uniqueness constraint. A uniqueness constraint assumes that only one correct match exists between stereo images. Some algorithms ignore repetitive objects and omit them in the depth map. We present a method that does not employ a uniqueness constraint, but rather determines whether an object is repetitive and then solves the matching problem by finding a unique object that is in close proximity to the object.

Paper Nr: 150
Title:

A RELATIONAL DISTANCE-BASED FRAMEWORK FOR HIERARCHICAL IMAGE UNDERSTANDING

Authors:

Laura Antanas, Martijn van Otterlo, José Oramas, Tinne Tuytelaars and Luc De Raedt

Abstract: Understanding images in terms of hierarchical and logical structures is crucial for many semantic tasks, including image retrieval, scene understanding and robot vision. This paper combines compositional hierarchies, qualitative spatial relations, relational instance-based learning and robust feature extraction in one framework. For each layer in the hierarchy, substructures in the images are detected, classified and then employed one layer up the hierarchy to obtain higher-level semantic structures, by making use of qualitative spatial relations. The approach is applied to street view images. We employ a four-layer hierarchy in which subsequently corners, windows and doors, and individual houses are detected.

Paper Nr: 169
Title:

PITCH-SENSITIVE COMPONENTS EMERGE FROM HIERARCHICAL SPARSE CODING OF NATURAL SOUNDS

Authors:

Engin Bumbacher and Vivienne Ming

Abstract: The neural basis of pitch perception, our subjective sense of the tone of a sound, has been a great ongoing debates in neuroscience.Variants of the two classic theories - spectral Place theory and temporal Timing theory - continue to continue to drive new experiments and debates (Shamma, 2004). Here we approach the question of pitch by applying a theoretical model based on the statistics of natural sounds. Motivated by gist research (Oliva and Torralba, 2006), we extended the nonlinear hierarchical generative model developed by Karklin et al. (Karklin and Lewicki, 2003) with a parallel gist pathway. The basic model encodes higher-order structure in natural sounds capturing variations in the underlying probability distribution. The secondary pathway provides a fast biasing of the model’s inference process based on the coarse spectrotemporal structures of sound stimuli on broader timescales. Adapting our extended model to speech demonstrates that the learned code describes a more detailed and broader range of statistical regularities that reflect abstract properties of sound such as harmonics and pitch than models without the gist pathway. The spectrotemporal modulation characteristics of the learned code are better matched to the modulation spectrum of speech signals than alternate models, and its higher-level coefficients capture information which not only effectively cluster related speech signals but also describe smooth transitions over time, encoding the temporal structure of speech signals. Finally, we find that the model produces a type of pitch-related density components which combine temporal and spectral qualities.

Paper Nr: 176
Title:

DOWNSCALING AEROSOL OPTICAL THICKNESS TO 1 KM2 SPATIAL RESOLUTION USING SUPPORT VECTOR REGRESSION REPLIED ON DOMAIN KNOWLEDGE

Authors:

Thi Nhat Thanh Nguyen, Simone Mantovani, Piero Campalani and Gian Piero Limone

Abstract: Processing of data recorded by MODIS sensors on board the polar orbiting satellite Terra and Aqua usually provides Aerosol Optical Thickness maps at a coarse spatial resolution. It is appropriate for applications of air pollution monitoring at the global scale but not adequate enough for monitoring at local scales. Different from the traditional approach based on physical algorithms to downscale the spatial resolution, in this article, we propose a methodology to derive AOT maps over land at 1 km2 of spatial resolution from MODIS data using support vector regression relied on domain knowledge. Experiments carried out on data recorded in three years over Europe areas show promising results on limited areas located around ground measurement sites where data are collected to make empirical data models as well as on large areas over satellite maps.

Paper Nr: 223
Title:

STACKED CONDITIONAL RANDOM FIELDS EXPLOITING STRUCTURAL CONSISTENCIES

Authors:

Peter Kluegl, Martin Toepfer, Florian Lemmerich, Andreas Hotho and Frank Puppe

Abstract: Conditional Random Fields (CRF) are popular methods for labeling unstructured or textual data. Like many machine learning approaches these undirected graphical models assume the instances to be independently distributed. However, in real world applications data is grouped in a natural way, e.g., by its creation context. The instances in each group often share additional consistencies in the structure of their information. This paper proposes a domain-independent method for exploiting these consistencies by combining two CRFs in a stacked learning framework. The approach incorporates three successive steps of inference: First, an initial CRF processes single instances as usual. Next, we apply rule learning collectively on all labeled outputs of one context to acquire descriptions of its specific properties. Finally, we utilize these descriptions as dynamic and high quality features in an additional (stacked) CRF. The presented approach is evaluated with a real-world dataset for the segmentation of references and achieves a significant reduction of the labeling error.

Paper Nr: 224
Title:

MUSIC GENRE CLASSIFICATION BASED ON DYNAMICAL MODELS

Authors:

Alberto García-Durán, Jerónimo Arenas-García, Darío García-García and Emilio Parrado-Hernández

Abstract: This paper studies several alternatives to extract dynamical features from hidden Markov Models (HMMs) that are meaningful for music genre supervised classification. Songs are modelled using a three scale approach: a first stage of short term (milliseconds) features, followed by two layers of dynamical models: a multivariate AR that provides mid term (seconds) features for each song followed by an HMM stage that captures long term (song) features shared among similar songs. We study from an empirical point of view which features are relevant for the genre classification task. Experiments on a database including pieces of heavy metal, punk, classical and reggae music illustrate the advantages of each set of features.

Short Papers
Paper Nr: 3
Title:

AN EMPIRICAL COMPARISON OF LABEL PREDICTION ALGORITHMS ON AUTOMATICALLY INFERRED NETWORKS

Authors:

Omar Ali, Giovanni Zappella, Tijl De Bie and Nello Cristianini

Abstract: The task of predicting the label of a network node, based on the labels of the remaining nodes, is an area of growing interest in machine learning, as various types of data are naturally represented as nodes in a graph. As an increasing number of methods and approaches are proposed to solve this task, the problem of comparing their performance becomes of key importance. In this paper we present an extensive experimental comparison of 15 different methods, on 15 different labelled-networks, as well as releasing all datasets and source code. In addition, we release a further set of networks that were not used in this study (as not all benchmarked methods could manage very large datasets). Besides the release of data, protocols and algorithms, the key contribution of this study is that in each of the 225 combinations we tested, the best performance—both in accuracy and running time—was achieved by the same algorithm: Online Majority Vote. This is also one of the simplest methods to implement.

Paper Nr: 8
Title:

IDENTIFYING DIAGNOSTIC EXPERTS - Measuring the Antecedents to Pattern Recognition

Authors:

Thomas Loveday, Mark Wiggins, Marino Festa and David Schell

Abstract: Medical expertise is typically denoted on the basis of experience, but this approach appears to lack validity and reliability. The present study investigated an innovative assessment of diagnostic expertise in medicine. This approach was developed from evidence that expert performance develops following the acquisition of cue associations in memory, which facilitates diagnostic pattern-recognition. Four distinct tasks were developed, for which the judicious extraction and selection of environmental cues may be advantageous. Across the tasks, performance clustered into two levels, reflecting competent and expert performance. These clusters were only weakly correlated with traditional methods of identifying domain experts, such as years of experience. The significance of this outcome is discussed in relation to training, evaluation and assessment.

Paper Nr: 11
Title:

ON APPLICATIONS OF SEQUENTIAL MULTI-VIEW DENSE RECONSTRUCTION FROM AERIAL IMAGES

Authors:

Dimitri Bulatov, Peter Wernerus and Hermann Gross

Abstract: Because of an increasing need and a rapid progress in the development of (unmanned) aerial vehicles and optical sensors that can be mounted onboard of these sensor platforms, there is also a considerable progress in 3D analysis of air- and UAV-borne video sequences. This work presents a robust method for multi-camera dense reconstruction as well as two important applications: creation of dense point clouds with precise 3D coordinates and, in the case of videos with Nadir perspective, a context-based method for urban terrain modeling. This method, which represents the main contribution of this work, includes automatic generation of digital terrain models (DTM), extraction of building outlines, modeling and texturing roof surfaces. A simple interactive method for vegetation segmentation is described as well.

Paper Nr: 12
Title:

THE COMBINATION OF HMAX AND HOGS IN AN ATTENTION GUIDED FRAMEWORK FOR OBJECT LOCALIZATION

Authors:

Tobias Brosch and Heiko Neumann

Abstract: Object detection and localization is a challenging task. Among several approaches, more recently hierarchical methods of feature-based object recognition have been developed and demonstrated high-end performance measures. Inspired by the knowledge about the architecture and function of the primate visual system, the computational HMAX model has been proposed. At the same time robust visual object recognition was proposed using feature distributions, e.g. histograms of oriented gradients (HOGs). Since both models build upon an edge representation of the input image, the question arises, whether one kind of approach might be superior to the other. Introducing a new biologically inspired attention steered processing framework, we demonstrate that the combination of both approaches gains the best results.

Paper Nr: 13
Title:

ESTIMATING PLANAR STRUCTURE IN SINGLE IMAGES BY LEARNING FROM EXAMPLES

Authors:

Osian Haines and Andrew Calway

Abstract: Outdoor urban scenes typically contain many planar surfaces, which are useful for tasks such as scene reconstruction, object recognition, and navigation, especially when only a single image is available. In such situations the lack of 3D information makes finding planes difficult; but motivated by how humans use their prior knowledge to interpret new scenes with ease, we develop a method which learns from a set of training examples, in order to identify planar image regions and estimate their orientation. Because it does not rely explicitly on rectangular structures or the assumption of a ‘Manhattan world’, our method can generalise to a variety of outdoor environments. From only one image, our method reliably distinguishes planes from non-planes, and estimates their orientation accurately; this is fast and efficient, with application to a real-time system in mind.

Paper Nr: 24
Title:

DYNAMICALLY MIXING DYNAMIC LINEAR MODELS WITH APPLICATIONS IN FINANCE

Authors:

Kevin R. Keane and Jason J. Corso

Abstract: Time varying model parameters offer tremendous flexibility while requiring more sophisticated learning methods. We discuss on-line estimation of time varying DLM parameters by means of a dynamic mixture model composed of constant parameter DLMs. For time series with low signal-to-noise ratios, we propose a novel method of constructing model priors. We calculate model likelihoods by comparing forecast distributions with observed values. We utilize computationally efficient moment matching Gaussians to approximate exact mixtures of path dependent posterior densities. The effectiveness of our approach is illustrated by extracting insightful time varying parameters for an ETF returns model in a period spanning the 2008 financial crisis. We conclude by demonstrating the superior performance of time varying mixture models against constant parameter DLMs in a statistical arbitrage application.

Paper Nr: 45
Title:

PERFORMANCE EVALUATION OF FEATURE DETECTION FOR LOCAL OPTICAL FLOW TRACKING

Authors:

Tobias Senst, Brigitte Unger, Ivo Keller and Thomas Sikora

Abstract: Due to its high computational efficiency the Kanade Lucas Tomasi feature tracker is still widely accepted and a utilized method to compute sparse motion fields or trajectories in video sequences. This method is made up of a Good Feature To Track feature detection and a pyramidal Lucas Kanade feature tracking algorithm. It is well known that the Good Feature To Track takes into account the Aperture Problem, but it does not consider the Generalized Aperture Problem. In this paper we want to provide an evaluation of a set of alternative feature detection methods. These methods are taken from feature matching techniques like FAST, SIFT and MSER. The evaluation is based on the Middlebury dataset and performed by using an improved pyramidal Lucas Kanade method, called RLOF feature tracker. To compare the results of the feature detector and RLOF pair, we propose a methodology based on accuracy, efficiency and covering measurements.

Paper Nr: 46
Title:

ON THE HUMAN POSE RECOVERY BASED ON A SINGLE VIEW

Authors:

Sébastien Piérard and Marc Van Droogenbroeck

Abstract: Estimating the pose of the observed person is crucial for a large variety of applications including home entertainment, man-machine interaction, video surveillance, etc. Often, only a single side view is available, but authors claim that it is possible to derive the pose despite that humans evolve in a 3D environment. In addition, to decrease the sensitivity to color and texture, it is preferable to rely only on the silhouette to recover the pose. Under these conditions, we show that there is an intrinsic limitation: at least two poses correspond to the observed silhouette. We discuss this intrinsic limitation in details in this short paper. To our knowledge, this issue has been overlooked by authors in the past. We observe that this limitation has an impact on the way previous reported results should be interpreted, and it has clearly to be taken into account for designing new methods.

Paper Nr: 48
Title:

TRACKING PLANAR-TEXTURED OBJECTS - On the Way to Transport Objects in Packaging Industry by Throwing and Catching

Authors:

Naeem Akhter

Abstract: In manufacturing systems transportation of objects can be optimized by throwing and catching them mechanically between work stations. There is a need to track thrown objects using visual sensors. Up to now only ball-shaped objects were tracked under controlled environment, where no orientation had to be considered. This work extends the task of object tracking to cuboid textured objects considering industrial environment. Indeed, tracking objects with respect to the robotics tasks to be achieved in a not too restricted environment remains an open issue. Thus, this work deals with efficient, flexible, and robust estimation of the object’s pose.

Paper Nr: 49
Title:

PHONEME-TO-VISEME MAPPING FOR VISUAL SPEECH RECOGNITION

Authors:

Luca Cappelletta and Naomi Harte

Abstract: Phonemes are the standard modelling unit in HMM-based continuous speech recognition systems. Visemes are the equivalent unit in the visual domain, but there is less agreement on precisely what visemes are, or how many to model on the visual side in audio-visual speech recognition systems. This paper compares the use of 5 viseme maps in a continuous speech recognition task. The focus of the study is visual-only recognition to examine the choice of viseme map. All the maps are based on the phoneme-to-viseme approach, created either using a linguistic method or a data driven method. DCT, PCA and optical flow are used to derive the visual features. The best visual-only recognition on the VidTIMIT database is achieved using a linguistically motivated viseme set. These initial experiments demonstrate that the choice of visual unit requires more careful attention in audio-visual speech recognition system development.

Paper Nr: 57
Title:

SIFT-BASED CAMERA LOCALIZATION USING REFERENCE OBJECTS FOR APPLICATION IN MULTI-CAMERA ENVIRONMENTS AND ROBOTICS

Authors:

Hanno Jaspers, Boris Schauerte and Gernot A. Fink

Abstract: In this contribution, we present a unified approach to improve the localization and the perception of a robot in a new environment by using already installed cameras. Using our approach we are able to localize arbitrary cameras in multi-camera environments while automatically extending the camera network in an online, unattended, real-time way. This way, all cameras can be used to improve the perception of the scene, and additional cameras can be added in real-time, e.g., to remove blind spots. To this end, we use the Scale-invariant feature transform (SIFT) and at least one arbitrary known-size reference object to enable camera localization. Then we apply non-linear optimization of the relative pose estimate and we use it to iteratively calibrate the camera network as well as to localize arbitrary cameras, e.g. of mobile phones or robots, inside a multi-camera environment. We performed an evaluation on synthetic as well as real data to demonstrate the applicability of the proposed approach.

Paper Nr: 64
Title:

A GENERALIZATION OF NEGATIVE NORM MODELS IN THE DISCRETE SETTING - Application to Stripe Denoising

Authors:

Jérôme Fehrenbach, Pierre Weiss and Corinne Lorenzo

Abstract: Starting with a book of Y.Meyer in 2001, negative norm models attracted the attention of the imaging community in the last decade. Despite numerous works, these norms seem to have provided only luckwarm results in practical applications. In this work, we propose a framework and an algorithm to remove stationary noise from images. This algorithm has numerous practical applications and we show it on 3D data from a newborn microscope called SPIM. We also show that this model generalizes Meyer’s model and its successors in the discrete setting and allows to interpret them in a Bayesian framework. It sheds a new light on these models and allows to pick them according to some a priori knowledge on the texture statistics. Further results are available on our webpage at http://www.math.univ-toulouse.fr/~weiss/PagePublications.html.

Paper Nr: 71
Title:

ROBUST FACE RECOGNITION USING WAVELET AND DCT BASED LIGHTING NORMALIZATION, AND SHIFTING-MEAN LDA

Authors:

I. Gede Pasek Suta Wijaya, Keiichi Uchimura, Gou Koutaki and Cuicui Zhang

Abstract: This paper presents an integration of Wavelet and Discrete Cosine Transform (DCT) based lighting normalization, and shifting-mean Linear Discriminant Analysis (LDA) based face classifiers for face recognition. The aims are to provide robust recognition rate against large face variability due to lighting variations and to avoid retraining problem of the classical LDA for incremental data. In addition, the compact holistic features is employed for dimensional reduction of the raw face image. From the experimental results, the proposed method gives sufficient and robust achievement in terms of recognition rate and requires short computational time.

Paper Nr: 72
Title:

A SYMBOLIC APPROACH FOR CLASSIFICATION OF MOVING VEHICLES IN TRAFFIC VIDEOS

Authors:

D. S. Guru, Elham Dallalzadeh and S. Manjunath

Abstract: In this paper, a symbolic approach is proposed to classify moving vehicles in traffic videos. A corner-based tracking method is presented to track and detect moving vehicles. We propose to overlap the boundary curves of each vehicle while tracking it in sequence of frames to reconstruct a complete boundary shape of the vehicle. The reconstructed boundary shape is normalized and then a set of efficient shape features are extracted. The extracted shape features are used to form interval-valued feature vector representation of vehicles. Vehicles are categorized into 4 different types of vehicle classes using a symbolic similarity measure. To corroborate the efficacy of the proposed method, experiment is conducted on 21,239 frames of roadway traffic videos taken in an uncontrolled environment during day time. The proposed method has 95.16% classification accuracy. Moreover, experiments reveal that the proposed method can be well adopted for on-line classification of moving vehicles as it is based on a simple matching scheme.

Paper Nr: 78
Title:

EFFICIENT COMPUTATION OF VORONOI NEIGHBORS BASED ON POLYTOPE SEARCH IN PATTERN RECOGNITION

Authors:

Juan Mendez and Javier Lorenzo

Abstract: Some algorithms in Pattern Recognition and Machine Learning as neighborhood-based classification and dataset condensation can be improved with the use of Voronoi tessellation. The paper shows the weakness of some existing algorithms of tessellation to deal with high dimensional datasets. The use of linear programming can improve the tessellation procedures by focusing in Voronoi adjacency. It will be shown that the adjacency test based on linear programming is a version of the polytope search. However, the polytope search procedure provides more information than a simple Boolean test. The paper proposes a strategy to use the additional information contained in the basis of the linear programming algorithm to obtain other tests. The theoretical results are applied to tessellate several random datasets, and also for much-used datasets in Machine Learning repositories.

Paper Nr: 84
Title:

NIGHT–TIME OUTDOOR SURVEILLANCE WITH MOBILE CAMERAS

Authors:

Ferran Diego, Georgios Evangelidis and Joan Serrat

Abstract: This paper addresses the problem of video surveillance by mobile cameras. We present a method that allows online change detection in night–time outdoor surveillance. Because of the camera movement, background frames are not available and must be ”localized“ in former sequences and registered with the current frames. To this end, we propose a Frame Localization And Registration (FLAR) approach that solves the problem efficiently. Frames of former sequences define a database which is queried by current frames in turn. To quickly retrieve nearest neighbors, database is indexed through a visual dictionary method based on the SURF descriptor. Furthermore, the frame localization is benefited by a temporal filter that exploits the temporal coherence of videos. Next, the recently proposed ECC alignment scheme is used to spatially register the synchronized frames. Finally, change detection methods apply to aligned frames in order to mark suspicious areas. Experiments with real night sequences recorded by in-vehicle cameras demonstrate the performance of the proposed method and verify its efficiency and effectiveness against other methods.

Paper Nr: 87
Title:

MEANINGFUL THICKNESS DETECTION ON POLYGONAL CURVE

Authors:

Bertrand Kerautret, Jacques-Olivier Lachaud and Mouhammad Said

Abstract: The notion of meaningful scale was recently introduced to detect the amount of noise present along a digital contour. It relies on the asymptotic properties of the maximal digital straight segment primitive. Even though very useful, the method is restricted to digital contour data and is not able to process other types of geometric data like disconnected set of points. In this work, we propose a solution to overcome this limitation. It exploits another primitive called the Blurred Segment which controls the straight segment recognition precision of disconnected sets of points. The resulting noise detection provides precise results and is also simpler to implement. A first application of contour smoothing demonstrates the efficiency of the proposed method. The algorithms can also be tested online.

Paper Nr: 89
Title:

FORECASTING US EQUITY LIQUIDITY SEASONALITY

Authors:

Aistis Raudys

Abstract: This paper considers liquidity seasonality patterns in US stock markets. We analyse weekly, monthly and annual liquidity seasonality across a set of NYSE and NASDAQ stocks. As a liquidity proxy we use trading volume. The paper demonstrates that long-term liquidity seasonality patterns exist and that incorporating seasonal information into forecasting models can noticeably benefit accuracy. In our research we create empirical non-parametric models to estimate future liquidity using simple averaging, robust regression, neural networks and k-nearest-neighbour techniques. To ensure the reliability of the conclusions we perform 100 experiments with long time frames using random splits of data into training and validation sets. By performing evaluations with six prediction methods we find that incorporating liquidity patterns benefits accuracy. More accurate forecasts could be used by market participants to reduce trading costs.

Paper Nr: 93
Title:

VEHICLE CLASSIFICATION USING EVOLUTIONARY FORESTS

Authors:

Murray Evans, Jonathan N. Boyle and James Ferryman

Abstract: Forests of decision trees are a popular tool for classification applications. This paper presents an approach to evolving the forest classifier, reducing the time spent designing the optimal tree depth and forest size. This is applied to the task of vehicle classification for purposes of verification against databases at security checkpoints, or accumulation of road usage statistics. The evolutionary approach to building the forest classifier is shown to out-perform a more typically grown forest and a baseline neural-network classifier for the vehicle classification task.

Paper Nr: 105
Title:

HOW TO SELECT USEFUL HAND SHAPES FOR HAND GESTURE RECOGNITION SYSTEM

Authors:

Atsushi Shimada, Takayoshi Yamashita and Rin-ichiro Tanguchi

Abstract: This paper discusses hand shapes for Human Computer Interface. Usually, a hand gesture based Human Computer Interface is developed by human centered design concept. A system designer or developer tends to select hand shapes by himself/herself without verifying practical effectiveness from the standpoint of system aspect. Instead, a methodology of training and recognition of hand shapes is often discussed. On the other hand, this paper listens to system’s voice; which hand shape is easy to be recognized, which is easy to be confused and so on. Actually, 37 kinds of tentative hand shapes were investigated from the viewpoint of system-friendly hand shape. Based on the result, a supporting system was developed for a system designer, which helps to find appropriate hand shapes which satisfy both “user-friendly” and “system-friendly” demand.

Paper Nr: 113
Title:

MODELLING A FOREGROUND FOR BACKGROUND SUBTRACTION FROM IMAGES - Probability Distribution of Pixel Positions based on Weighted Intensity Differences

Authors:

Suil Son, Young-Woon Cha and Suk I. Yoo

Abstract: To overcome a false detection problem caused by dynamic textures in background subtraction problems, a new modelling approach is suggested. While traditional background subtraction approaches model the background, an indirect method, to detect foreground objects, the approach described here models the foreground directly. The foreground model is given by the probability distribution of pixel positions in terms of sums of weighted intensity differences for each pixel position between all previous images and a new image. The combination of the weighting and the summing of the intensity differences produces a number of desirable effects. For instance, each position in the new image which has consistently large differences will have a high foreground probability value; each position having consistently small differences will have a low probability value; and positions having small differences for most of the previous images but large differences for a few of the previous images due to dynamic textures or noises will have medium probability values. The final distribution of the foreground position is computed by Kernel density estimation incorporating the neighboring pixel differences, and foreground objects are then identified by the probability value of this distribution. The performance of the suggested approach is then illustrated with two classes of problems and compared to other conventional approaches.

Paper Nr: 119
Title:

A NEW PROBLEM IN FACE IMAGE ANALYSIS - Finding Kinship Clues for Siblings Pairs

Authors:

A. Bottino, M. De Simone, A. Laurentini and T. Vieira

Abstract: Human face conveys to other human beings, and potentially to computers, much information such as identity, emotional states, intentions, age and attractiveness. Among this information there are kinship clues. Face kinship signals, as well as the human capabilities of capturing them, are studied by psychologist and sociologists. In this paper we present a new research aimed at analyzing, with image processing/pattern analysis techniques, facial images for detecting objective elements of similarity between siblings. To this end, we have constructed a database of high quality pictures of pairs of siblings, shot in controlled conditions, including frontal, profile, expressionless and smiling face images. A first analysis of the database has been performed using a commercial identity recognition software. Then, for discriminating siblings, we combined eigenfaces, SVM and a feature selection algorithm, obtaining a recognition accuracy close to that of a human rating panel.

Paper Nr: 127
Title:

3D BENDING OF SURFACES AND VOLUMES WITH AN APPLICATION TO BRAIN TORQUE MODELING

Authors:

Antonietta Pepe and Jussi Tohka

Abstract: In this work, we propose a novel space deformation model for local bending of 3D volumes and surfaces. The model can be easily controlled through accommodation of a few intuitive parameters. Experiments on volumes, parametric surfaces, and polygonal surfaces show that our method has increased modeling capabilities when compared to the previous space deformation methods for local bending. We apply this new, more flexible model for space bending to model human brain asymmetry. In particular, we develop an image processing pipeline for automatic generation of a set of realistic 3D brain magnetic resonance (MR) images for which the asymmetry is known. This dataset can be used for the quantitative validation of voxel and surface based methods for studying brain shape asymmetry. The pipeline encompasses a realistic modeling of the anatomical rightward bending of the inter-hemispheric fissure in human brain.

Paper Nr: 130
Title:

EFFICIENT GAIT-BASED GENDER CLASSIFICATION THROUGH FEATURE SELECTION

Authors:

Raúl Martín- Félez, Javier Ortells, Ramón A. Mollineda and J. Salvador Sánchez

Abstract: Apart from human recognition, gait has lately become a promising biometric feature also useful for prediction of gender. One of the most popular methods to represent gait is the well-known Gait Energy Image (GEI), which conducts to a high-dimensional Euclidean space where many features are irrelevant. In this paper, the problem of selecting the most relevant GEI features for gender classification is addressed. In particular, an ANOVA-based algorithm is used to measure the discriminative power of each GEI pixel. Then, a binary mask is built from the few most significant pixels in order to project a given GEI onto a reduced feature pattern. Experiments over two large gait databases show that this method leads to similar recognition rates to those of using the complete GEI, but with a drastic dimensionality reduction. As a result, a much more efficient gender classification model regarding both computing time and storage requirements is obtained.

Paper Nr: 136
Title:

MONOCULAR EGOMOTION ESTIMATION BASED ON IMAGE MATCHING

Authors:

Diego Cheda, Daniel Ponsa and Antonio Manuel López

Abstract: In this paper, we propose a novel method for computing the egomotion of a monocular camera mounted on a vehicle based on the matching of distant regions in consecutive frames. Our approach takes advantage of the fact that the image projection of a plane can provide information about the camera motion. Instead of tracking points between frames, we track distant regions in the scene because they behave as an infinity plane. As a consequence of tracking this infinity plane, we obtain an image geometric transformation (more precisely, an infinity homography) relating two consecutive frames. This transformation is actually capturing the camera rotation, since the effect produced by the translation can be neglected at long distances. Then, we can compute the camera rotation as the result of the previously estimated infinity homography. After that, rotation can be canceled from images, just leading to a translation explaining the motion between two frames. Experiments on real image sequences show that our approach reaches higher accuracy w.r.t. state-of-the-art methods.

Paper Nr: 140
Title:

IMPROVING FEATURE LEVEL LIKELIHOODS USING CLOUD FEATURES

Authors:

Heydar Maboudi Afkham, Stefan Carlsson and Josephine Sullivan

Abstract: The performance of many computer vision methods depends on the quality of the local features extracted from the images. For most methods the local features are extracted independently of the task and they remain constant through the whole process. To make features more dynamic and give models a choice in the features they can use, this work introduces a set of intermediate features referred as cloud features. These features take advantage of part-based models at the feature level by combining each extracted local feature with its close by local feature creating a cloud of different representations for each local features. These representations capture the local variations around the local feature. At classification time, the best possible representation is pulled out of the cloud and used in the calculations. This selection is done based on several latent variables encoded within the cloud features. The goal of this paper is to test how the cloud features can improve the feature level likelihoods. The focus of the experiments of this paper is on feature level inference and showing how replacing single features with equivalent cloud features improves the likelihoods obtained from them. The experiments of this paper are conducted on several classes of MSRCv1 dataset.

Paper Nr: 143
Title:

DEPTH CAMERA TECHNOLOGY COMPARISON AND PERFORMANCE EVALUATION

Authors:

Benjamin Langmann, Klaus Hartmann and Otmar Loffeld

Abstract: How good are cheap depth cameras, namely the Microsoft Kinect, compared to state of the art Time-of-Flight depth cameras? In this paper several depth cameras of different types were put to the test on a variety of tasks in order to judge their respective performance and to find out their weaknesses. We will concentrate on the common area of applications for which both types are specified, i.e. near field indoor scenes. The characteristics and limitations of the different technologies as well as the concrete hardware implementations are discussed and evaluated with a set of experimental setups. Especially, the noise level and the axial and angular resolutions are compared. Additionally, refined formulas to generate depth values based on the raw measurements of the Kinect are presented.

Paper Nr: 156
Title:

EMBEDDED FEATURE SELECTION FOR SPAM AND PHISHING FILTERING USING SUPPORT VECTOR MACHINES

Authors:

Sebastián Maldonado and Gastón L'Huillier

Abstract: Today, the Internet is full of harmful and wasteful elements, such as phishing and spam messages, which must be properly classified before reaching end-users. This issue has attracted the pattern recognition community’s attention and motivated to determine which strategies achieve best classification results. Several methods use as many features as content-based properties the data set have, which leads to a high dimensional classification problem. In this context, this paper presents a feature selection approach that simultaneously determines a nonlinear classification function with minimal error and minimizes the number of features by penalizing their use in the dual formulation of binary Support Vector Machines (SVM). The method optimizes the width of an anisotropic RBF Kernel via successive gradient descent steps, eliminating features that have low relevance for the model. Experiments with two real-world Spam and Phishing data sets demonstrate that our approach accomplishes the best performance compared to well-known feature selection methods using consistently a small number of features.

Paper Nr: 157
Title:

ROTATIONAL INVARIANCE AT FIXATION POINTS - Experiments using Human Gaze Data

Authors:

Johannes Steffen, Christian Hentschel, Afra'a Ahmad Alyosef, Klaus Toennies and Andreas Nuernberger

Abstract: An important aspect in machine vision concerns the extraction of meaningful patterns at salient image regions. Invariance w.r.t. affine transformations has usually been claimed to be a crucial attribute of these regions. While continuing research on the human visual cortex has suggested the correctness of these assumptions at least in later stages of vision, only lately the availability of accurate and cheap eye tracking devices has offered the possibility to provide empirical evidence to these claims. We present an experimental setting that is qualified to analyse various assumptions on human gaze target properties. The proposed setting aims at reducing high-level influence on the fixation process as much as possible. As a proof of concept we present results for the assumption human fixation targeting is rotational invariant. Even though high-level aspects could not be completely suppressed, we were able to detect and analyse this relation in the gaze data. It was found that there is a significant correlation between fixated regions within stimuli over different orientations.

Paper Nr: 161
Title:

AUTOMATIC ESTIMATION OF MULTIPLE MOTION FIELDS USING OBJECT TRAJECTORIES AND OPTICAL FLOW

Authors:

Manya V. Afonso, Jorge S. Marques and Jacinto C. Nascimento

Abstract: Multiple motion fields are an efficient way of summarising the movement of objects in a scene and allow an automatic classification of objects activities in the scene. However, their estimation relies on some kind of supervised learning e.g., using manually edited trajectories. This paper proposes an automatic method for the estimation of multiple motion fields. The proposed algorithm detects multiple moving objects and their velocities in a video sequence using optical flow. This leads to a sequence of centroids and corresponding velocity vectors. A matching algorithm is then applied to group the centroids into trajectories, each of them describing the movement of an object in the scene. The paper shows that motion fields can be reliably estimated from the detected trajectories leading to a fully automatic procedure for the estimation of multiple motion fields.

Paper Nr: 166
Title:

OPTICAL FLOW ESTIMATION WITH CONFIDENCE MEASURES FOR SUPER-RESOLUTION BASED ON RECURSIVE ROBUST TOTAL LEAST SQUARES

Authors:

Tobias Schuchert and Fabian Oser

Abstract: In this paper we propose a novel optical flow estimation method accompanied by confidence measures. Our main goal is fast and highly accurate motion estimation in regions where information is available and a confidence measure which identifies these regions. Therefore we extend the structure tensor method to robust recursive total least squares (RRTLS) and run it on a GPU for real-time processing. Based on a coarse-to-fine framework we propagate not only the motion estimates to finer scales but also the covariance matrices, which may be used as confidence measures. Experiments on synthetic data show the benefits of our approach. We applied the RRTLS framework to a real-time super-resolution method for deforming objects which incorporates the confidence measures and demonstrates that propagating the covariances through the pyramid improves super-resolution results.

Paper Nr: 183
Title:

SHIFT AND ROTATION INVARIANT IRIS FEATURE EXTRACTION BASED ON NON-SUBSAMPLED CONTOURLET TRANSFORM AND GLCM

Authors:

Sirvan Khalighi, Parisa Tirdad, Fatemeh Pak and Urbano Nunes

Abstract: A new feature extraction method for iris recognition in non-subsampled contourlet transform (NSCT) domain is proposed. To extract the features a two-level NSCT, which is a shift-invariant transform, and a rotation-invariant gray level co-occurrence matrix (GLCM) with 3 different orientations are applied on both spatial image and NSCT frequency subbands. The extracted feature set is transformed and normalized to reduce the effect of extreme values in the feature matrix. A set of significant features are selected by using the minimal redundancy and maximal relevance (mRMR) algorithm. Finally the selected feature set is classified using support vector machines (SVMs). The classification results using leave one out cross-validation (LOOCV) on the CASIA iris database, Ver.1 and Ver.4 show that the proposed method performs at the state-of-the art in the field of iris recognition.

Paper Nr: 187
Title:

PROBABILISTIC AFFECTIVE MODEL WITH PROBABILISTIC LATENT SEMANTIC ANALYSIS FOR PREDICTING HUMAN AFFECT

Authors:

Yunhee Shin and Eun Yi Kim

Abstract: As affective computing becomes increasingly important, it becomes necessary to retrieve and process images according to human affect or preference. However, judging such affective qualities of images is highly subjective task. Despite a lack of firm rules, certain features in images are expected to be more related than certain others. Among the various visual features that can be associated with affective classes, this work focuses on color composition, which can have the most influence on affective level. In this paper, we suggest predicting certain affective features included in an image using color compositions that constitutes the image. For this, we propose a Probabilistic Affective Model (PAM) to automatically estimate the probabilities that an image is related to certain affective classes. To construct the affective space, we consider learning-based mapping approach. The proposed learning-based method considered probabilistic latent semantic analysis (PLSA) which is used in the field of textual analysis, which can catch the latent context between terms and documents. We adapt this latent model to mapping approach to better suit in our problem. Potential applications include social network service, re-ranking, canonical image selection and design of web page interfaces.

Paper Nr: 201
Title:

MELANOSOME TRACKING BY BAYES THEOREM AND ESTIMATION OF MOVABLE REGION

Authors:

Toshiaki Okabe and Kazuhiro Hotta

Abstract: This paper proposes a melanosome tracking method using Bayes theorem and estimation of movable region of melanosome candidates. Melanosomes in intracellular images are tracked manually now to investigate the cause of disease, and automatic tracking method is desired. Since there are little automatic recognition methods for intracellular images, we can not know which features and classifiers are effective for them. Thus, we try to develop the melanosome tracking using Bayes theorem of melanosome candidates detected by Scale-Invariant Feature Transform (SIFT). However, SIFT can not detect the center of melanosome because melanosome is too small in images. Therefore, SIFT detector is adopted after image size is enlarged by Lanczos resampling. However, there are still many melanosome candidates. Thus, we estimate the movable region of the target melanosome in next frame and eliminate melanosome candidates. After the posterior probability of each candidate is computed by Bayes theorem, and the melanosome with the maximum probability is tracked. Experimental results using the melanosome images of normal and Griscelli syndrome show the effectiveness of our method.

Paper Nr: 206
Title:

STREAM VOLUME PREDICTION IN TWITTER WITH ARTIFICIAL NEURAL NETWORKS

Authors:

Gabriela Dominguez, Juan Zamora, Miguel Guevara, Héctor Allende and Rodrigo Salas

Abstract: Twitter is one of the most important social network, where extracting useful information is of paramount importance to many application areas. Many works to date have tried to mine this information by taking the network structure, language itself or even by searching for a pattern in the words employed by the users. Anyway, a simple idea that might be useful for every challenging mining task - and that at out knowledge has not been tackled yet - consists of predicting the amount of messages (stream volume) that will be emitted in some specific time span. In this work, by using almost 180k messages collected in a period of one week, a preliminary analysis of the temporal structure of the stream volume in Twitter is made. The expected contribution consists of a model based on artificial neural networks to predict the amount of posts in a specific time window, which regards the past history and the daily behavior of the network in terms of the emission rate of the message stream.

Paper Nr: 210
Title:

SHAPE REPRESENTATION AND A MORPHING SCHEME TO SUPPORT FLAPPING WING RESEARCH

Authors:

Mohammad Sharif Khan and Tapabrata Ray

Abstract: Wing geometry is one of the most important factors that affects the performance of a flapping wing. The shape of insect wings and their nature of flapping varies across insect species. In order to gain an in-depth understanding of flapping flight with an aim to identify optimal wing shapes, there is a need for an universal and flexible shape representation scheme that is amenable to optimization. The paper presents a methodology to represent boundaries of insect wings which can be subsequently morphed via an optimization algorithm. The shapes are represented using B-splines, wherein the control points representing the shapes are repaired and subsequently evolved within an optimization framework. Twelve insect-wing shapes have been used to test the performance of the proposed method in the context of shape matching.

Posters
Paper Nr: 4
Title:

HUMAN ACTION RECOGNITION USING CONTINUOUS HMMS AND HOG/HOF SILHOUETTE REPRESENTATION

Authors:

Mohamed Ibn Khedher, Mounim A. El-Yacoubi and Bernadette Dorizzi

Abstract: This paper presents an alternative to the mainstream approach of STIP-based SVM recognition for human recognition. First, it studies whether or not whole silhouette representation by Histogram-of-Oriented-Gradients (HOG) or Histogram-of-Optical-Flow (HOF) descriptors is more discriminated when compared to sparse spatio-temporal interest points (STIPs). Second, it investigates whether explicitly modeling the temporal order of features using continuous HMMs outperforms the standard Bag-of-Words (BoW) representation that overlooks such an order. When both whole silhouette representation and temporal order modeling are combined, a significant improvement is shown on the Weizmann database over STIP-based SVM recognition.

Paper Nr: 7
Title:

PUBSEARCH - A Hierarchical Heuristic Scheme for Ranking Academic Search Results

Authors:

Emmanouil Amolochitis, Ioannis T. Christou and Zheng-Hua Tan

Abstract: In this paper we present PubSearch, a meta-search engine system for academic publications. We have designed a ranking algorithm consisting of a hierarchical set of heuristic models including term frequency, depreciated citation count and a graph-based score for associations among paper index terms. We used our algorithm to re-rank the default search results produced by online digital libraries such as ACM Portal in response to specific user-submitted queries. The experimental results show that the ranking algorithm used by our system can provide a more relevant ranking scheme compared to ACM Portal.

Paper Nr: 19
Title:

VISUAL NAVIGATION FOR THE BLIND - Path and Obstacle Detection

Authors:

J. José, J. M. H. du Buf and J. M. F. Rodrigues

Abstract: We present a real-time vision system to assist blind and visually impaired persons. This system complements the white cane, and it can be used both indoor and outdoor. It detects borders of paths and corridors, obstacles within the borders, and it provides guidance for centering and obstacle avoidance. Typical obstacles are backpacks, trash cans, trees, light poles, holes, branches, stones and other objects at a distance of 2 to 5 meters from the camera position. Walkable paths are detected by edges and an adapted Hough transform. Obstacles are detected by a combination of three algorithms: zero crossings of derivatives, histograms of binary edges, and Laws’ texture masks.

Paper Nr: 27
Title:

LANDMARK EXTRACTION FROM LEAVES WITH PALMATE VENATION - Application to Grape

Authors:

Raffi Enficiaud and Sofiène Mouine

Abstract: The growing interest of Content Base Image Retrieval techniques in the context of plant identification requires the development of appropriate features. A considerable amount of information about the taxonomic identity of a plant is contained in its leaves, and most of the botanical expertise uses jointly the contour and the venation network. The current work focuses principally on the extraction of the venation network, the base and secondary landmarks of leaves with uncluttered background, assuming only their structure as palmate. Morphological operators are used to extract a first approximation of the venation network, which is then filtered by a voting scheme and reconstructed using a connected component like algorithm. The base point and the primary veins are then extracted with an accuracy of 100%, which allows identification of the lobes and the measurement their relative length.

Paper Nr: 42
Title:

DECISION TREE INDUCTION FROM COUNTEREXAMPLES

Authors:

Nicolas Cebron, Fabian Richter and Rainer Lienhart

Abstract: While it is well accepted in human learning to learn from counterexamples or mistakes, classic machine learning algorithms still focus only on correctly labeled training examples.We replace this rigid paradigm by using complementary probabilities to describe the probability that a certain class does not occur. Based on the complementary probabilities, we design a decision tree algorithm that learns from counterexamples. In a classification problem with K classes, K 􀀀1 counterexamples correspond to one correctly labeled training example. We demonstrate that even when only a partial amount of counterexamples is available, we can still obtain good performance.

Paper Nr: 51
Title:

DETECTION OF LANE DEPARTURE ON HIGH-SPEED ROADS

Authors:

David Hanwell and Majid Mirmehdi

Abstract: We present a system for detecting and tracking the lanes of high-speed roads, in order to warn the driver of accidental lane departures. The proposed method introduces a novel variant of the classic Hough transform, better equipped to detect and locate linear road markings with a common vanishing point. This is combined with a simple model of the lane and an Extended Kalman Filter to make detection and tracking more robust. This allows detection of lane changes, resistant to visual interference by traffic and irrelevant road markings.

Paper Nr: 52
Title:

DEAL EFFECT CURVE AND PROMOTIONAL MODELS - Using Machine Learning and Bootstrap Resampling Test

Authors:

Cristina Soguero-Ruiz, Francisco Javier Gimeno-Blanes, Inmaculada Mora-Jiménez, María Pilar Martínez-Ruiz and José Luis Rojo-Álvarez

Abstract: Promotional sales have become in recent years a paramount issue in the marketing strategies of many companies, specially in the current economic situation. Empirical models of consumer promotional behavior, mostly based on machine learning methods, are becoming more usual than theoretical models, given the complexity of the promotional interactions and the availability of electronic recordings. However, the performance description and comparison among promotion models are usually made in terms of absolute and empirical values, which is a limited handling of the information. Here we first propose to use a simple nonparametric statistical tool, the paired bootstrap resampling, for establishing clear cut-off test based comparisons among methods for machine learning based promotional models, by simply taking into account the estimated statistical distribution of the actual risk. The method is used to determine the existence of actual statistically significant differences in the performance of different machine design issues for multilayer perceptron based marketing models, in a real database of everyday goods (milk products). Our results show that paired bootstrap resampling is a simple and effective procedure for promotional modeling using machine learning techniques.

Paper Nr: 60
Title:

ACTIVE STEREO-MATCHING FOR ONE-SHOT DENSE RECONSTRUCTION

Authors:

Sergio Fernandez, Josep Forest and Joaquim Salvi

Abstract: Stereo-vision in computer vision represents an important field for 3D reconstruction. Real time dense reconstruction, however, is only achieved for high textured surfaces in passive stereo-matching. In this work an active stereo-matching approach is proposed. A projected pattern is used to artificially increase the texture of the measuring object, thus enabling dense reconstruction for one-shot stereo techniques. Results show that the accuracy is similar to other active techniques, while dense reconstruction is obtained.

Paper Nr: 74
Title:

SMART METER - Artificial Neural Network for Disaggregation of Electrical Appliances

Authors:

Dirk Benyoucef, Thomas Bier and Philipp Klein

Abstract: Goal of that paper is to show a possibility for the disaggregation of electrical appliances in the load curve of residential buildings. The advantage is that the measurement system is at a central point in the household. So the installation effort decrease. For the disaggregation of the appliances out of the load curve, an approach for the development of classification algorithms is presented. One method for the classification of appliances is to use Artificial Neural Network. This idea is the main part of that paper. It is shown a method, to classify one kind of appliances. At the end, the first relsults and the next steps are presented. The disaggregation of the appliances is part of a research project at the University of Furtwangen.

Paper Nr: 79
Title:

DETECTING 'YELLOW STAIN' IN WOOD USING SPECTRAL METHODS

Authors:

Gerald McGunnigle

Abstract: This paper deals with the detection of ‘yellow stain’ in wood samples using colour. We describe an investigation into the spectral properties of the stain and use the findings to design a detection system. We found that infected regions invariably differed from healthy regions in the 400nm to 450nm region of the spectrum. We developed a system based on an economical RGB camera and an optimised light source. The source consists of two narrow bands: one in the blue channel of the camera where the discriminative information is held, and one in the red channel that acts as a normalisation factor to remove the effect of the natural patterning of the wood. A simple classifier was used with the red and blue channels of the camera and produced results that agreed with our client’s subjective judgement.

Paper Nr: 80
Title:

DETECTION AND RECOGNITION OF SUBPIXEL TARGETS WITH HYPOTHESES DEPENDENT BACKGROUND POWER

Authors:

Victor Golikov and Olga Lebedeva

Abstract: We consider the problem of detecting and recognizing the subpixel targets in sea background when the background power may be different under the null hypothesis – where it is assumed to be known – and the alternative multiple hypotheses. This situation occurs when the presence of the target triggers a decrease in the background power (subpixel targets). We extend the formulation of the Matched Subspace Detector (MSD) to the case where the background power is only known under the null hypothesis using the generalized likelihood ratio test (GLRT) for the multiple hypotheses case. The obtained multiple hypotheses test is based on the Modified MSD test (MMSD). We discuss the difference between the two detection and recognition systems: based on the MSD and MMSD tests. Numerical simulations attest to the validity of the performance analysis.

Paper Nr: 96
Title:

INDOOR EVALUATION OF CROP ROWAND GRID DETECTION - System for an Automated Transplanter

Authors:

J. Reumers, F. De Smedt, J. Anthonis, H. Ramon and T. Goedemé

Abstract: If crops with considerable spacing can be arranged in a precise rectangular pattern, mechanical weeding can become an alternative to chemical methods by enabling treatment in two perpendicular directions. Realisation of such a pattern requires innovation of seedling transplanters. A computer vision-based sensing system was developed for detecting a transplanters posture relative to the crop. A method for indoor evaluation of the systems accuracy is proposed, using an experimental cart on a set of rails to control the vehicle’s posture. The method was successful in evaluating the estimates of the lateral offset and the heading angle, but the reliability of the validation values for the longitudinal distance is limited. Tests results showed that the maximum error on the measurements of the lateral offset is 1.5cm and that the standard deviation is smaller than 0.6cm. The maximum error on the angle measurements is 2:1º. The standard deviation of the error is smaller than 0:65º. The standard deviation of the error on the estimates of the longitudinal distance is typically 0.7cm.

Paper Nr: 102
Title:

FUZZY TEMPLATES FOR PAIR-WISE MULTI-CLASS CLASSIFICATION

Authors:

Rimantas Kybartas

Abstract: In this paper, a pair-wise classifier fusion by Fuzzy Templates is analyzed. Single layer perceptron and support vectors are used as pair-wise classifiers. Comments on weakness and strength of such fusion method are presented. The obtained results show, that in some cases such an approach could be competitive or even outperform other pair-wise classifier fusion methods.

Paper Nr: 124
Title:

DETECTION AND RECOVERY OF OCCLUDED FACE IMAGES BASED ON CORRELATION BETWEEN PIXELS

Authors:

Ji-eun Lee and Nojun Kwak

Abstract: In this paper, we propose a method to detect and recover the occluded parts of face images using the correlation between pairs of pixels. In the training stage, correlation coefficients between every pairs of pixels are calculated using the occlusion-free training face images. Once a new face image is shown, the occluded area is detected and recovered using correlation coefficients obtained in the training stage. We compare the performance of the proposed method with the conventional method based on PCA. The results show that the proposed method detects and recovers occluded area with much smaller noises than the conventional PCA based method.

Paper Nr: 144
Title:

SHAPE RECOGNITION USING THE LEAST SQUARES APPROXIMATION

Authors:

Nacéra Laiche and Slimane Larabi

Abstract: This paper represents a novel algorithm to represent and recognize two dimensional curve based on its convex hull and the Least-Squared modeling. It combines the advantages of the property of the convex hulls that are particularly suitable for affine matching as they are affine invariant and the geometric properties of a contour that make it more or less identifiable. The description scheme and the similarity measure developed take into consideration technique for shape similarity. According to this method, the contours are extracted and decomposed into portions of curves. Each portion curve is approximated by some explicit curve using the Least Squares approximation. The obtained cubic curves are normalized in order to make the method invariant to scale change. Finally the resulting curves are used to compare and to compute similarity between shapes in images database using the Hausdorff distance. The proposed algorithm has been tested and its performance is found favourable as compared to other matching techniques.

Paper Nr: 149
Title:

ICOUNTER - Development of an Optical Readout Method fot Mechanical Counters

Authors:

Dirk Benyoucef, Pirmin Held and Thomas Bier

Abstract: Mechanical counters are still very common in electricity, water and gas meters. Automatic readout of the dial count without modifying the mechanics of the counter is only possible using expensive image processing methods. Therefore the topic of this report is a new method for automatically reading out the counter values without the need of additional mechanical or parallel electronic parts inside the counter. Instead the different reflection properties of the different digits are measured and evaluated. This is done using only simple electronic parts and a microcontroller. In the first part of the paper the hardware for measuring the reflection values is presented. A model of this hardware with special emphasis on the influences of the environment is discussed in the next part. Following this, two classification methods, for distinguishing the digits are analyzed. For showing the properties of the new readout system measurements and simulations are given in the end.

Paper Nr: 164
Title:

USER INDEPENDENT SYSTEM FOR RECOGNITION OF HAND POSTURES USED IN SIGN LANGUAGE

Authors:

Dahmani Djamila, Benchikh Soumia and Slimane Larabi

Abstract: A new signer independent method of recognition of hand postures of sign language alphabet is presented in this paper. We propose a new geometric hand postures features derived from the convex hull enclosing the hand’s shape. These features are combined with the discrete orthogonal Tchebichef moments, and the Hu moments. The Tchebichef moments are applied on the external and internal edges of the hand’s shape. Experiments, based on two different hand posture data sets, show that our method is robust at recognizing hand postures independent of the person performing them. The system obtains a good recognition rates, and also performs well compared to other hand user independent posture recognition systems.

Paper Nr: 174
Title:

DECOMPOSITION OF MULTIMODAL DATA FOR AFFORDANCE-BASED IDENTIFICATION OF POTENTIAL GRASPS

Authors:

Daniel Dornbusch, Robert Haschke, Stefan Menzel and Heiko Wersing

Abstract: In this paper, we apply standard decomposition approaches to the problem of finding local correlations in multi-modal and high- dimensional grasping data, particularly to correlate the local shape of cup-like objects to their associated local grasp configurations. We compare the capability of several decomposition methods to establish these task-relevant, inter-modal correlations and indicate how they can be exploited to find potential contact points and hand postures for novel, though similar, objects.

Paper Nr: 180
Title:

UNDERSTANDING TOA AND TDOA NETWORK CALIBRATION USING FAR FIELD APPROXIMATION AS INITIAL ESTIMATE

Authors:

Yubin Kuang, Erik Ask, Simon Burgess and Kalle Åström

Abstract: This paper presents a study of the so called far field approximation to the problem of determining both the direction to a number of transmittors and the relative motion of a single antenna using relative distance measurements. The same problem is present in calibration of microphone and wifi-transmittor arrays. In the far field approximation we assume that the relative motion of the antenna is small in comparison to the distances to the base stations. The problem can be solved uniquely with at least three motions of the antenna and at least six real or virtual transmittors. The failure modes of the problem is determined to be (i) when the antenna motion is planar or (ii) when the transmittor directions lie on a cone. We also study to what extent the solution can be obtained in these degenerate configurations. The solution algorithm for the minimal case can be extended to the overdetermined case in a straightforward manner. We also implement and test algorithms for non-linear optimization of the residuals. In experiments we explore how sensitive the calibration is with respect to different degrees of far field approximations of the transmittors and with respect to noise in the data.

Paper Nr: 203
Title:

GREEDY APPROACH FOR DOCUMENT CLUSTERING

Authors:

Lim Choen Choi and Soon Cheol Park

Abstract: A Greedy Algorithm for Document Clustering (Greedy Clustering) is proposed in this paper. Various cluster validity indices (DB, CH, SD, AS) are used to find the most appropriate optimization function for Greedy Clustering. The clustering algorithms are tested and compared on Reuter-21578. The results show that AS Index shows the best performance and the fastest running time among cluster indices in various experiments. Also Greedy Clustering with AS Index has 15~20% better performance than traditional clustering algorithms (K-means, Group Average).