[pt] Nesta tese, propomos novas arquiteturas para redes neurais profundas utlizando métodos de atenção e álgebra multilinear para aumentar seu desempenho. Também exploramos convoluções em grafos e suas particularidades. Nos concentramos aqui em problemas relacionados à estimativa de pose em tempo real. A estimativa de pose é um problema desafiador em visão computacional com muitas aplicações reais em áreas como realidade aumentada, realidade virtual, animação por computador e reconstrução de cenas 3D. Normalmente, o problema a ser abordado envolve estimar a pose humana 2D ou 3D, ou seja, as partes do corpo de pessoas em imagens ou vídeos, bem como seu posicionamento e estrutura. Diveros trabalhos buscam atingir alta precisão usando arquiteturas baseadas em redes neurais de convolução convencionais; no entanto, erros causados por oclusão e motion blur não são incomuns, e ainda esses modelos são computacionalmente pesados para aplicações em tempo real. Exploramos diferentes arquiteturas para melhorar o tempo de processamento destas redes e, como resultado, propomos dois novos modelos de rede neural para estimativa de pose 2D e 3D. Também apresentamos uma nova arquitetura para redes de atenção em grafos chamada de atenção em grafos semânticos. / [en] This thesis proposes new architectures for deep neural networks with attention enhancement and multilinear algebra methods to increase their performance. We also explore graph convolutions and their particularities. We focus here on the problems related to real-time pose estimation. Pose estimation is a challenging problem in computer vision with many real applications in areas including augmented reality, virtual reality, computer animation, and 3D scene reconstruction. Usually, the problem to be addressed involves estimating the 2D and 3D human pose, i.e., the anatomical keypoints or body parts of persons in images or videos. Several papers propose approaches to achieve high accuracy using architectures based on conventional convolution neural networks; however, mistakes caused by occlusion and motion blur are not uncommon, and those models are computationally very intensive for real-time applications. We explore different architectures to improve processing time, and, as a result, we propose two novel neural network models for 2D and 3D pose estimation. We also introduce a new architecture for Graph attention networks called Semantic Graph Attention.

This report examines the problem of real-time head pose estimation in low-resolution football footage. A method is presented for inferring the head pose using a combination of footage and knowledge of the locations of the football and players. An ensemble of randomized ferns is compared with a support vector machine for processing the footage, while a support vector machine performs pattern recognition on the location data. Combining the two sources of information outperforms either in isolation. The location of the football turns out to be an important piece of information. / QC 20100707 / Capturing and Visualizing Large scale Human Action (ACTVIS)

No description available.

Bien qu'ayant suscité des recherches depuis 30 ans, le problème de la reconnaissance de visages en contexte de vidéosurveillance, sachant qu'une seule image par individu est disponible pour l'enrôlement, n'est pas encore résolu. Dans ce contexte, les deux dés les plus diciles à relever consistent à développer des algorithmes robustes aux variations d'illumination et aux variations de pose. De plus, il y a aussi une contrainte forte sur la complexité en temps et en occupation mémoire des algorithmes à mettre en oeuvre dans de tels systèmes. Le travail développé dans cette thèse apporte plusieurs avancées innovantes dans ce contexte de reconnaissance faciale en vidéosurveillance. Premièrement, une méthode de normalisation des variations d'illumination visant à simuler les performances de la rétine est proposée en tant que pré-traitement des images faciales. Deuxièmement, nous proposons un nouveau descripteur appelé POEM (Patterns of Oriented Edge Magnitudes) destiné à représenter les structures locales d'une image. Ce descripteur est discriminant, robuste aux variations extérieures (variations de pose, d'illumination, d'expression, d'âge que l'on rencontre souvent avec les visages). Troisièmement, un modèle statistique de reconnaissance de visages en conditions de pose variables, centré sur une modélisation de la manière dont l'apparence du visage évolue lorsque le point de vue varie, est proposé. Enn, une nouvelle approche visant à modéliser les relations spatiales entre les composantes du visage est présentée. A l'exception de la dernière approche, tous les algorithmes proposés sont très rapides à calculer et sont donc adaptés à la contrainte de traitement temps réel des systèmes de vidéosurveillance.

Cette thèse s'inscrit dans le cadre de PILE, un projet médical d'analyse du regard, des gestes, et des productions vocales d'enfants en bas âge. Dans ce contexte, nous avons conçu et développé des méthodes de détermination de l'orientation de la tête, pierre angulaire des systèmes d'estimation de la direction du regard. D'un point de vue méthodologique, nous avons proposé BISAR (Boosted Input Selection Algorithm for Regression), une méthode de sélection de caractéristiques adaptée aux problèmes de régression. Elle consiste à sélectionner itérativement les entrées d'un réseau de neurones incrémental. Chaque entrée est associée à un descripteur sélectionné à l'aide d'un critère original qui mesure la dépendance fonctionnelle entre un descripteur et les valeurs à prédire. La complémentarité des descripteurs est assurée par un processus de boosting qui modifie, à chaque itération, la distribution des poids associés aux exemples d'apprentissage. Cet algorithme a été validé expérimentalement au travers de deux méthodes d'estimation de la pose de la tête. La première approche apprend directement la relation entre l'apparence d'un visage et sa pose. La seconde aligne un modèle de visage dans une image, puis estime géométriquement l'orientation de ce modèle. Le processus d'alignement repose sur une fonction de coût qui évalue la qualité de l'alignement. Cette fonction est apprise par BISAR à partir d'exemples de modèles plus ou moins bien alignés. Les évaluations de ces méthodes ont donné des résultats équivalents ou supérieurs aux méthodes de l'état de l'art sur différentes bases présentant de fortes variations de pose, d'identité, d'illumination et de conditions de prise de vues.

This thesis presents new methods in two closely related areas of computer vision: human pose estimation, and gesture recognition in videos. In human pose estimation, we show that random forests can be used to estimate human pose in monocular videos. To this end, we propose a co-segmentation algorithm for segmenting humans out of videos, and an evaluator that predicts whether the estimated poses are correct or not. We further extend this pose estimator to new domains (with a transfer learning approach), and enhance its predictions by predicting the joint positions sequentially (rather than independently) in an image, and using temporal information in the videos (rather than predicting the poses from a single frame). Finally, we go beyond random forests, and show that convolutional neural networks can be used to estimate human pose even more accurately and efficiently. We propose two new convolutional neural network architectures, and show how optical flow can be employed in convolutional nets to further improve the predictions. In gesture recognition, we explore the idea of using weak supervision to learn gestures. We show that we can learn sign language automatically from signed TV broadcasts with subtitles by letting algorithms 'watch' the TV broadcasts and 'match' the signs with the subtitles. We further show that if even a small amount of strong supervision is available (as there is for sign language, in the form of sign language video dictionaries), this strong supervision can be combined with weak supervision to learn even better models.

Usage of real objects as links between real and virtual information is one key aspect in augmented reality. A central issue to achieve this link is the estimation of the visuospatial information of the observed object, or in other words, estimating the object pose. Different objects can have different behaviors when used for interaction. This not only encompasses changes in position, but also folding or deformations. Traditional researches in the area solve those pose estimation problems using different approaches, depending on the type of the object. Additionally, some researches are based only on positional information of observed feature points, simplifying the object information. In this work, we explore the pose estimation of different objects by gathering more information from the observed feature points, and obtaining the local poses of such points, which are not explored in other researches. We apply this local pose estimation idea in two different capturing scenarios, reaching two novel approaches of pose estimation: one based on RGB-D cameras, and another based on RGB and machine learning methods. In the RGB-D based approach, we use the feature point orientation and near surface to obtain its normal; then, find the local 6 degrees-of-freedom (DoF) pose. This approach gives us not only the rigid object pose, but also the approximated pose of deformed objects. On the other hand, our RGB based approach explores machine learning with local appearance changes. Unlike other RGB based works, we replace the complex non-linear systems solvers with a fast and robust method, reaching local rotation of the observed feature points, as well as, full 6 DoF rigid object pose with dramatically lower real-time calculation demands. Both approaches show us that gathering local poses can bring information for the pose estimation of different types of objects. / O uso de objetos reais como meio de conexão entre informações reais e virtuais é um aspecto chave dentro da realidade aumentada. Uma questão central para tal conexão é a estimativa de informações visuo-espaciais do objeto, ou em outras palavras, a detecção da pose do objeto. Diferentes objetos podem ter diferentes comportamentos quando utilizados em interações. Não somente incluindo a mudança de posição, mas também sendo dobradas ou deformadas. Pesquisas tradicionais solucionam tais problemas de detecção usando diferentes abordagens, dependendo do tipo de objeto. Adicionalmente, algumas pesquisas se baseiam somente na informação posicional dos pontos de interesse, simplificando a informação do objeto. Neste trabalho, a detecção de pose de diferente objetos é explorada coletando-se mais informações dos pontos de interesse observados e, por sua vez, obtendo as poses locais de tais pontos, poses que não são exploradas em outras pesquisas. Este conceito da detecção de pose locais é aplicada em dois ambientes de capturas, estendendo-se em duas abordagens inovadoras: uma baseada em câmeras RGB-D, e outra baseada em câmeras RGB e métodos de aprendizado de maquinas. Na abordagem baseada em RGB-D, a orientação e superfície ao redor do ponto de interesse são utilizadas para obter a normal do ponto. Através de tais informações a pose local é obtida. Esta abordagem não só permite a obtenção de poses de objetos rígidos, mas também a pose aproximada de objetos deformáveis. Por outro lado, a abordagem baseada em RGB explora o aprendizado de máquina aplicado em alterações das aparências locais. Diferentemente de outros trabalhos baseados em câmeras RGB, esta abordagem substitui solucionadores não lineares complexos com um método rápido e robusto, permitindo a obtenção de rotações locais dos pontos de interesse, assim como, a pose completa (com 6 graus-de-liberdade) de objetos rígidos, com uma demanda computacional muito menor para cálculos em tempo-real. Ambas as abordagens mostram que a coleta de poses locais podem gerar informações para a detecção de poses de diferentes tipos de objetos.

This thesis investigates the task of estimating the pose of a hand interacting with an object from a depth image. The main contribution of this thesis is the development of our discriminative one-shot hand-object pose estimation system. To the best of our knowledge, this is the first attempt at a one-shot hand-object pose estimation system. It is a two stage system consisting of convolutional neural networks. The first stage segments the object out of the hand from the depth image. This hand-minus-object depth image is combined with the original input depth image to form a 2-channel image for use in the second stage, pose estimation. We show that using this 2-channel image produces better pose estimation performance than a single stage pose estimation system taking just the input depth map as input. We also believe that we are amongst the first to research hand-object segmentation. We use fully convolutional neural networks to perform hand-object segmentation from a depth image. We show that this is a superior approach to random decision forests for this task. Datasets were created to train our hand-object pose estimator stage and hand-object segmentation stage. The hand-object pose labels were estimated semi-automatically with a combined manual annotation and generative approach. The segmentation labels were inferred automatically with colour thresholding. To the best of our knowledge, there were no public datasets for these two tasks when we were developing our system. These datasets have been or are in the process of being publicly released.

Usage of real objects as links between real and virtual information is one key aspect in augmented reality. A central issue to achieve this link is the estimation of the visuospatial information of the observed object, or in other words, estimating the object pose. Different objects can have different behaviors when used for interaction. This not only encompasses changes in position, but also folding or deformations. Traditional researches in the area solve those pose estimation problems using different approaches, depending on the type of the object. Additionally, some researches are based only on positional information of observed feature points, simplifying the object information. In this work, we explore the pose estimation of different objects by gathering more information from the observed feature points, and obtaining the local poses of such points, which are not explored in other researches. We apply this local pose estimation idea in two different capturing scenarios, reaching two novel approaches of pose estimation: one based on RGB-D cameras, and another based on RGB and machine learning methods. In the RGB-D based approach, we use the feature point orientation and near surface to obtain its normal; then, find the local 6 degrees-of-freedom (DoF) pose. This approach gives us not only the rigid object pose, but also the approximated pose of deformed objects. On the other hand, our RGB based approach explores machine learning with local appearance changes. Unlike other RGB based works, we replace the complex non-linear systems solvers with a fast and robust method, reaching local rotation of the observed feature points, as well as, full 6 DoF rigid object pose with dramatically lower real-time calculation demands. Both approaches show us that gathering local poses can bring information for the pose estimation of different types of objects. / O uso de objetos reais como meio de conexão entre informações reais e virtuais é um aspecto chave dentro da realidade aumentada. Uma questão central para tal conexão é a estimativa de informações visuo-espaciais do objeto, ou em outras palavras, a detecção da pose do objeto. Diferentes objetos podem ter diferentes comportamentos quando utilizados em interações. Não somente incluindo a mudança de posição, mas também sendo dobradas ou deformadas. Pesquisas tradicionais solucionam tais problemas de detecção usando diferentes abordagens, dependendo do tipo de objeto. Adicionalmente, algumas pesquisas se baseiam somente na informação posicional dos pontos de interesse, simplificando a informação do objeto. Neste trabalho, a detecção de pose de diferente objetos é explorada coletando-se mais informações dos pontos de interesse observados e, por sua vez, obtendo as poses locais de tais pontos, poses que não são exploradas em outras pesquisas. Este conceito da detecção de pose locais é aplicada em dois ambientes de capturas, estendendo-se em duas abordagens inovadoras: uma baseada em câmeras RGB-D, e outra baseada em câmeras RGB e métodos de aprendizado de maquinas. Na abordagem baseada em RGB-D, a orientação e superfície ao redor do ponto de interesse são utilizadas para obter a normal do ponto. Através de tais informações a pose local é obtida. Esta abordagem não só permite a obtenção de poses de objetos rígidos, mas também a pose aproximada de objetos deformáveis. Por outro lado, a abordagem baseada em RGB explora o aprendizado de máquina aplicado em alterações das aparências locais. Diferentemente de outros trabalhos baseados em câmeras RGB, esta abordagem substitui solucionadores não lineares complexos com um método rápido e robusto, permitindo a obtenção de rotações locais dos pontos de interesse, assim como, a pose completa (com 6 graus-de-liberdade) de objetos rígidos, com uma demanda computacional muito menor para cálculos em tempo-real. Ambas as abordagens mostram que a coleta de poses locais podem gerar informações para a detecção de poses de diferentes tipos de objetos.

In order for assistive robots to collaborate effectively with humans for completing everyday tasks, they must be endowed with the ability to effectively perceive scenes and more importantly, recognize human intentions. As a result, we present in this dissertation a novel scene-dependent human-robot collaborative system capable of recognizing and learning human intentions based on scene objects, the actions that can be performed on them, and human interaction history. The aim of this system is to reduce the amount of human interactions necessary for communicating tasks to a robot. Accordingly, the system is partitioned into scene understanding and intention recognition modules. For scene understanding, the system is responsible for segmenting objects from captured RGB-D data, determining their positions and orientations in space, and acquiring their category labels. This information is fed into our intention recognition component where the most likely object and action pair that the user desires is determined. Our contributions to the state of the art are manifold. We propose an intention recognition framework that is appropriate for persons with limited physical capabilities, whereby we do not observe human physical actions for inferring intentions as is commonplace, but rather we only observe the scene. At the core of this framework is our novel probabilistic graphical model formulation entitled Object-Action Intention Networks. These networks are undirected graphical models where the nodes are comprised of object, action, and object feature variables, and the links between them indicate some form of direct probabilistic interaction. This setup, in tandem with a recursive Bayesian learning paradigm, enables our system to adapt to a user's preferences. We also propose an algorithm for the rapid estimation of position and orientation values of scene objects from single-view 3D point cloud data using a multi-scale superquadric fitting approach. Additionally, we leverage recent advances in computer vision for an RGB-D object categorization procedure that balances discrimination and generalization as well as a depth segmentation procedure that acquires candidate objects from tabletops. We demonstrate the feasibility of the collaborative system presented herein by conducting evaluations on multiple scenes comprised of objects from 11 categories, along with 7 possible actions, and 36 possible intentions. We achieve approximately 81% reduction in interactions overall after learning despite changes to scene structure.