The aim of this thesis is to provide methods for 2D segmentation and 2D/3D tracking, that are both fast and robust to imperfect image information, as caused for example by occlusions, motion blur and cluttered background. We do this by combining high level shape information with simultaneous segmentation and tracking. We base our work on the assumption that the space of possible 2D object shapes can be either generated by projecting down known rigid 3D shapes or learned from 2D shape examples. We minimise the discrimination between statistical foreground and background appearance models with respect to the parameters governing the shape generative process (the 6 degree-of-freedom 3D pose of the 3D shape or the parameters of the learned space). The foreground region is delineated by the zero level set of a signed distance function, and we define an energy over this region and its immediate background surroundings based on pixel-wise posterior membership probabilities. We obtain the differentials of this energy with respect to the parameters governing shape and conduct searches for the correct shape using standard non-linear minimisation techniques. This methodology first leads to a novel rigid 3D object tracker. For a known 3D shape, our optimisation here aims to find the 3D pose that leads to the 2D projection that best segments a given image. We extend our approach to track multiple objects from multiple views and propose novel enhancements at the pixel level based on temporal consistency. Finally, owing to the per pixel nature of much of the algorithm, we support our theoretical approach with a real-time GPU based implementation. We next use our rigid 3D tracker in two applications: (i) a driver assistance system, where the tracker is augmented with 2D traffic sign detections, which, unlike previous work, allows for the relevance of the traffic signs to the driver to be gauged and (ii) a robust, real time 3D hand tracker that uses data from an off-the-shelf accelerometer and articulated pose classification results from a multiclass SVM classifier. Finally, we explore deformable 2D/3D object tracking. Unlike previous works, we use a non-linear and probabilistic dimensionality reduction, called Gaussian Process Latent Variable Models, to learn spaces of shape. Segmentation becomes a minimisation of an image-driven energy function in the learned space. We can represent both 2D and 3D shapes which we compress with Fourier-based transforms, to keep inference tractable. We extend this method by learning joint shape-parameter spaces, which, novel to the literature, enable simultaneous segmentation and generic parameter recovery. These can describe anything from 3D articulated pose to eye gaze. We also propose two novel extensions to standard GP-LVM: a method to explore the multimodality in the joint space efficiently, by learning a mapping from the latent space to a space that encodes the similarity between shapes and a method for obtaining faster convergence and greater accuracy by use of a hierarchy of latent embeddings.

Dans les systèmes d'aide à la décision, sont généralement à disposition des données numériques abondantes et éventuellement certaines connaissances contextuelles qualitatives, disponibles a priori ou fournies a posteriori par retour d'expérience. Les performances des approches de classification, en particulier spectrale, dépendent de l'intégration de ces connaissances dans leur conception. Les algorithmes de classification spectrale permettent de traiter la classification sous l'angle de coupes de graphe. Ils classent les données dans l'espace des vecteurs propres de la matrice Laplacienne du graphe. Cet espace est censé mieux révéler la présence de groupements naturels linéairement séparables. Dans ce travail, nous nous intéressons aux algorithmes intégrant des connaissances type contraintes de comparaison. L'espace spectral doit, dans ce cas, révéler la structuration en classes tout en respectant, autant que possible, les contraintes de comparaison. Nous présentons un état de l'art des approches spectrales semi-supervisées contraintes. Nous proposons un nouvel algorithme qui permet de générer un sous-espace de projection par optimisation d'un critère de multi-coupes normalisé avec ajustement des coefficients de pénalité dus aux contraintes. Les performances de l'algorithme sont mises en évidence sur différentes bases de données par comparaison à d'autres algorithmes de la littérature. Dans le cadre de la surveillance de l'écosystème marin, nous avons développé un système de classification automatique de cellules phytoplanctoniques, analysées par cytométrie en flux. Pour cela, nous avons proposé de mesurer les similarités entre cellules par comparaison élastique entre leurs signaux profils caractéristiques. / In the decision support systems, often, there a huge digital data and possibly some contextual knowledge available a priori or provided a posteriori by feedback. The performances of classification approaches, particularly spectral ones, depend on the integration of the domain knowledge in their design. Spectral classification algorithms address the problem of classification in terms of graph cuts. They classify the data in the eigenspace of the graph Laplacian matrix. The generated eigenspace may better reveal the presence of linearly separable data clusters. In this work, we are particularly interested in algorithms integrating pairwise constraints : constrained spectral clustering. The eigenspace may reveal the data structure while respecting the constraints. We present a state of the art approaches to constrained spectral clustering. We propose a new algorithm, which generates a subspace projection, by optimizing a criterion integrating both normalized multicut and penalties due to the constraints. The performances of the algorithms are demonstrated on different databases in comparison to other algorithms in the literature. As part of monitoring of the marine ecosystem, we developed a phytoplankton classification system, based on flow cytometric analysis. for this purpose, we proposed to characterize the phytoplanktonic cells by similarity measures using elastic comparison between their cytogram signals.

O reconhecimento de faces é uma área de pesquisa desafiadora que abre portas para a implementação de aplicações muito promissoras. Embora muitos algoritmos eficientes e robustos já tenham sido propostos, ainda restam vários desafios. Dentre os principais obstáculos a serem uperados, está a obtenção de uma representação robusta e compacta de faces que possibilite distinguir os indivíduos rapidamente. Visando abordar esse problema, foi realizado um estudo de técnicas de reconhecimento estatístico de padrões, principalmente na área de redução de dimensionalidade dos dados, além de uma revisão de métodos de reconhecimento de faces. Foi proposto (em colaboração com a pesquisadora Isabelle Bloch) um método de seleção de características que une um algoritmo de busca eficiente (métodos de busca seqüencial flutuante) com uma medida de distância entre conjuntos nebulosos (distância nebulosa baseada em tolerância). Essa medida de distância possui diversas vantagens, sendo possível considerar as diferentes tipicalidades de cada padrão dos conjuntos de modo a permitir a obtenção de bons resultados mesmo com conjuntos com sobreposição. Os resultados preliminares com dados sintéticos mostraram o caráter promissor dessa abordagem. Com o objetivo de verificar a eficiência de tal técnica com dados reais, foram efetuados testes com reconhecimento de pessoas usando imagens da região dos olhos. Nesse caso, em se tratando de um problema com mais de duas classes, nós propusemos uma nova função critério inspirada na distância supracitada. Além disso foi proposto (juntamente com o estudante de mestrado Rogério S. Feris) um esquema de reconhecimento a partir de seqüências de vídeo. Esse esquema inclui a utilização de um método eficiente de rastreamento de características faciais (Gabor Wavelet Networks) e o método proposto anteriormente para seleção de características. Dentro desse contexto, o trabalho desenvolvido nesta dissertação implementa uma parte dos módulos desse esquema. / Face recognition is an instigating research field that may lead to the development of many promising applications. Although many efficient and robust algorithms have been developed in this area, there are still many challenges to be overcome. In particular, a robust and compact face representation is still to be found, which would allow for quick classification of different individuals. In order to address this problem, we first studied pattern recognition techniques, especially regarding dimensionality reduction, followed by the main face recognition methods. We introduced a new feature selection approach in collaboration with the researcher Isabelle Bloch (TSI-ENST-Paris), that associates an efficient searching algorithm (sequential floating search methods), with a tolerance-based fuzzy distance. This distance measure presents some nice features for dealing with the tipicalities of each pattern in the sets, so that good results can be attained even when the sets are overlapping. Preliminary results with synthetic data have demonstrated that this method is quite promising. In order to verify the efficiency of this technique with real data, we applied it for improving the performance of a person recognition system based on eye images. Since this problem involves more than two classes, we also developed a new criterion function based on the above-mentioned distance. Moreover, we proposed (together with Rogério S. Feris) a system for person recognition based on video sequences. This mechanism includes the development of an efficient method for facial features tracking, in addition to our method for feature selection. In this context, the work presented here constitutes part of the proposed system.

Made available in DSpace on 2017-07-21T14:19:33Z (GMT). No. of bitstreams: 1 Juscelino Izidoro Oliveira.pdf: 622255 bytes, checksum: 54447b380bca4ea8e2360060669d5cff (MD5) Previous issue date: 2012-07-30 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior / Multivariate data analysis allows the researcher to verify the interaction among a lot of attributes that can influence the behavior of a response variable. That analysis uses models that can be induced from experimental data set. An important issue in the induction of multivariate regressors and classifers is the sample size, because this determines the reliability of the model for tasks of regression or classification of the response variable. This work approachs the sample size issue through the Theory of Probably Approximately Correct Learning, that comes from problems about machine learning for induction of models. Given the importance of agricultural modelling, this work shows two procedures to select variables. Variable Selection by Principal Component Analysis is an unsupervised procedure and allows the researcher to select the most relevant variables from the agricultural data by considering the variation in the data. Variable Selection by Supervised Principal Component Analysis is a supervised procedure and allows the researcher to perform the same process as in the previous procedure, but concentrating the focus of the selection over the variables with more influence in the behavior of the response variable. Both procedures allow the sample complexity informations to be explored in variable selection process. Those procedures were tested in five experiments, showing that the supervised procedure has allowed to induce models that produced better scores, by mean, than that models induced over variables selected by unsupervised procedure. Those experiments also allowed to verify that the variables selected by the unsupervised and supervised procedure showed reduced indices of multicolinearity. / A análise multivariada de dados permite verificar a interação de vários atributos que podem influenciar o comportamento de uma variável de resposta. Tal análise utiliza modelos que podem ser induzidos de conjuntos de dados experimentais. Um fator importante na indução de regressores e classificadores multivariados é o tamanho da amostra, pois, esta determina a contabilidade do modelo quando há a necessidade de se regredir ou classificar a variável de resposta. Este trabalho aborda a questão do tamanho da amostra por meio da Teoria do Aprendizado Provavelmente Aproximadamente Correto, oriundo de problemas sobre o aprendizado de máquina para a indução de modelos. Dada a importância da modelagem agrícola, este trabalho apresenta dois procedimentos para a seleção de variáveis. O procedimento de Seleção de Variáveis por Análise de Componentes Principais, que não é supervisionado e permite ao pesquisador de agricultura selecionar as variáveis mais relevantes de um conjunto de dados agrícolas considerando a variação contida nos dados. O procedimento de Seleção de Variáveis por Análise de Componentes Principais Supervisionado, que é supervisionado e permite realizar o mesmo processo do primeiro procedimento, mas concentrando-se apenas nas variáveis que possuem maior infuência no comportamento da variável de resposta. Ambos permitem que informações a respeito da complexidade da amostra sejam exploradas na seleção de variáveis. Os dois procedimentos foram avaliados em cinco experimentos, mostrando que o procedimento supervisionado permitiu, em média, induzir modelos que produziram melhores pontuações do que aqueles modelos gerados sobre as variáveis selecionadas pelo procedimento não supervisionado. Os experimentos também permitiram verificar que as variáveis selecionadas por ambos os procedimentos apresentavam índices reduzidos de multicolinaridade..

Over the past ten years, more than 4,000 genome-wide association studies (GWAS) have helped to shed light on the genetic architecture of complex traits and diseases. In recent years, phenotyping of the samples has often gone beyond single traits and it has become common to record multi- to high-dimensional phenotypes for individu- als. Whilst these rich datasets offer the potential to analyse complex trait structures and pleiotropic effects at a genome-wide level, novel analytic challenges arise. This thesis summarises my research into genetic associations for high-dimensional phen- otype data. First, I developed a novel and computationally efficient approach for multivari- ate analysis of high-dimensional phenotypes based on linear mixed models, com- bined with bootstrapping (LiMMBo). Both in simulation studies and on real data, I demonstrate the statistical validity of LiMMBo and that it can scale to hundreds of phenotypes. I show the gain in power of multivariate analyses for high-dimensional phenotypes compared to univariate approaches, and illustrate that LiMMBo allows for detecting pleiotropy in a large number of phenotypic traits. Aside from their computational challenges in GWAS, the true dimensionality of very high-dimensional phenotypes is often unknown and lies hidden in high-dimen- sional space. Retaining maximum power for association studies of such phenotype data relies on using an appropriate phenotype representation. I systematically ana- lysed twelve unsupervised dimensionality reduction methods based on their per- formance in finding a robust phenotype representation in simulated data of different structure and size. I propose a stability criteria for choosing low-dimensional phen- otype representations and demonstrate that stable phenotypes can recover genetic associations. Finally, I analysed genetic variants for associations to high-dimensional cardiac phenotypes based on MRI data from 1,500 healthy individuals. I used an unsuper- vised approach to extract a low-dimensional representation of cardiac wall thickness and conducted a GWAS on this representation. In addition, I investigated genetic associations to a trabeculation phenotype generated from a supervised feature ex- traction approach on the cardiac MRI data. In summary, this thesis highlights and overcomes some of the challenges in per- forming genetic association studies on high-dimensional phenotypes. It describes new approaches for phenotype processing, and genotype to phenotype mapping for high-dimensional datasets, as well as providing new insights in the genetic structure of cardiac morphology in humans.

O reconhecimento de faces é uma área de pesquisa desafiadora que abre portas para a implementação de aplicações muito promissoras. Embora muitos algoritmos eficientes e robustos já tenham sido propostos, ainda restam vários desafios. Dentre os principais obstáculos a serem uperados, está a obtenção de uma representação robusta e compacta de faces que possibilite distinguir os indivíduos rapidamente. Visando abordar esse problema, foi realizado um estudo de técnicas de reconhecimento estatístico de padrões, principalmente na área de redução de dimensionalidade dos dados, além de uma revisão de métodos de reconhecimento de faces. Foi proposto (em colaboração com a pesquisadora Isabelle Bloch) um método de seleção de características que une um algoritmo de busca eficiente (métodos de busca seqüencial flutuante) com uma medida de distância entre conjuntos nebulosos (distância nebulosa baseada em tolerância). Essa medida de distância possui diversas vantagens, sendo possível considerar as diferentes tipicalidades de cada padrão dos conjuntos de modo a permitir a obtenção de bons resultados mesmo com conjuntos com sobreposição. Os resultados preliminares com dados sintéticos mostraram o caráter promissor dessa abordagem. Com o objetivo de verificar a eficiência de tal técnica com dados reais, foram efetuados testes com reconhecimento de pessoas usando imagens da região dos olhos. Nesse caso, em se tratando de um problema com mais de duas classes, nós propusemos uma nova função critério inspirada na distância supracitada. Além disso foi proposto (juntamente com o estudante de mestrado Rogério S. Feris) um esquema de reconhecimento a partir de seqüências de vídeo. Esse esquema inclui a utilização de um método eficiente de rastreamento de características faciais (Gabor Wavelet Networks) e o método proposto anteriormente para seleção de características. Dentro desse contexto, o trabalho desenvolvido nesta dissertação implementa uma parte dos módulos desse esquema. / Face recognition is an instigating research field that may lead to the development of many promising applications. Although many efficient and robust algorithms have been developed in this area, there are still many challenges to be overcome. In particular, a robust and compact face representation is still to be found, which would allow for quick classification of different individuals. In order to address this problem, we first studied pattern recognition techniques, especially regarding dimensionality reduction, followed by the main face recognition methods. We introduced a new feature selection approach in collaboration with the researcher Isabelle Bloch (TSI-ENST-Paris), that associates an efficient searching algorithm (sequential floating search methods), with a tolerance-based fuzzy distance. This distance measure presents some nice features for dealing with the tipicalities of each pattern in the sets, so that good results can be attained even when the sets are overlapping. Preliminary results with synthetic data have demonstrated that this method is quite promising. In order to verify the efficiency of this technique with real data, we applied it for improving the performance of a person recognition system based on eye images. Since this problem involves more than two classes, we also developed a new criterion function based on the above-mentioned distance. Moreover, we proposed (together with Rogério S. Feris) a system for person recognition based on video sequences. This mechanism includes the development of an efficient method for facial features tracking, in addition to our method for feature selection. In this context, the work presented here constitutes part of the proposed system.

A method to detect relationships between disease susceptibility and multilocus genetic interactions is the Multifactor-Dimensionality Reduction (MDR) technique pioneered by Ritchie et al. (2001). Since its introduction, many extensions have been pursued to deal with non-binary outcomes and/or account for multiple interactions simultaneously. Studying the effects of multilocus genetic interactions on continuous traits (blood pressure, weight, etc.) is one case that MDR does not handle. Culverhouse et al. (2004) and Gui et al. (2013) proposed two different methods to analyze such a case. In their research, Gui et al. (2013) introduced the Quantitative Multifactor-Dimensionality Reduction (QMDR) that uses the overall average of response variable to classify individuals into risk groups. The classification mechanism may not be efficient under some circumstances, especially when the overall mean is close to some multilocus means. To address such difficulties, we propose a new algorithm, the Ordered Combinatorial Quantitative Multifactor-Dimensionality Reduction (OQMDR), that uses a series of testings, based on ascending order of multilocus means, to identify best interactions of different orders with risk patterns that minimize the prediction error. Ten-fold cross-validation is used to choose from among the resulting models. Regular permutations testings are used to assess the significance of the selected model. The assessment procedure is also modified by utilizing the Generalized Extreme-Value distribution to enhance the efficiency of the evaluation process. We presented results from a simulation study to illustrate the performance of the algorithm. The proposed algorithm is also applied to a genetic data set associated with Alzheimer's Disease.

One of the hardest challenges in the field of machine learning is to build agents, such as robotic assistants in homes and hospitals, that can autonomously learn new tasks that they were not pre-programmed to tackle, without the intervention of an engineer. Reinforcement learning (RL) and learning from demonstration (LfD) are popular approaches for task learning, but they are often ineffective in high-dimensional domains unless provided with either a great deal of problem-specific domain information or a carefully crafted representation of the state and dynamics of the world. Unfortunately, autonomous agents trying to learn new tasks usually do not have access to such domain information nor to an appropriate representation. We demonstrate that algorithms that focus, at each moment, on the relevant features of the state space can achieve significant speed-ups over previous reinforcement learning algorithms with respect to the number of state features in complex domains. To do so, we introduce and evaluate a family of attention focus algorithms. We show that these algorithms can reduce the dimensionality of complex domains, creating a compact representation of the state space with which satisficing policies can be learned efficiently. Our approach obtains exponential speed-ups with respect to the number of features considered when compared with table-based learning algorithms and polynomial speed-ups when compared with state-of-the-art function approximation RL algorithms such as LSPI or fitted Q-learning. Our attention focus algorithms are divided in two classes, depending on the source of the focus information they require. Attention focus from human demonstrations infers the features to focus on from a set of demonstrations from human teachers performing the task the agent must learn. We introduce two algorithms within this class. The first one, abstraction from demonstration (AfD), identifies features that can be safely ignored in the whole state space and builds a state-space abstraction where a satisficing policy can be learned efficiently. The second, automatic decomposition and abstraction from demonstration, goes one step further, using the demonstrations to identify a set of subtasks and to find an appropriate abstraction for each subtask found. The other class of algorithms we present, attention focus with a world model, does not require a set of human demonstrations. Instead, it extracts the attention focus information from an object-based model of the world together with the agent experience in performing the task. Within this class, we introduce object-focused Q-learning (OF-Q), at first with an assumption of object independence that is later removed to support domains where objects interact with each other. Finally, we show that both sources of focus information can be combined for further speed-ups.

In this thesis, I look at three problems with important applications in data processing. Incorporating side information, provided by the user or derived from data, is a main theme of each of these problems. <br /><br /> This thesis makes a number of contributions. The first is a technique for combining different embedding objectives, which is then exploited to incorporate side information expressed in terms of transformation invariants known to hold in the data. It also introduces two different ways of incorporating transformation invariants in order to make new similarity measures. Two algorithms are proposed which learn metrics based on different types of side information. These learned metrics can then be used in subsequent embedding methods. Finally, it introduces a manifold learning algorithm that is useful when applied to sequential decision problems. In this case we are given action labels in addition to data points. Actions in the manifold learned by this algorithm have meaningful representations in that they are represented as simple transformations.

In this thesis, I look at three problems with important applications in data processing. Incorporating side information, provided by the user or derived from data, is a main theme of each of these problems. <br /><br /> This thesis makes a number of contributions. The first is a technique for combining different embedding objectives, which is then exploited to incorporate side information expressed in terms of transformation invariants known to hold in the data. It also introduces two different ways of incorporating transformation invariants in order to make new similarity measures. Two algorithms are proposed which learn metrics based on different types of side information. These learned metrics can then be used in subsequent embedding methods. Finally, it introduces a manifold learning algorithm that is useful when applied to sequential decision problems. In this case we are given action labels in addition to data points. Actions in the manifold learned by this algorithm have meaningful representations in that they are represented as simple transformations.