Deep learning architecture for gesture recognition

Pós-graduação em Engenharia Mecânica - FEG

Methodologies for visual correspondence

Visual correspondence is a key computer vision task that aims at identifying projections of the same 3D point into images taken either from different viewpoints or at different time instances. This task has been the subject of intense research activities in the last years in scenarios such as object recognition, motion detection, stereo vision, pattern matching, image registration. The approaches proposed in literature typically aim at improving the state of the art by increasing the reliability, the accuracy or the computational efficiency of visual correspondence algorithms. The research work carried out during the Ph.D. course and presented in this dissertation deals with three specific visual correspondence problems: fast pattern matching, stereo correspondence and robust image matching. The dissertation presents original contributions to the theory of visual correspondence, as well as applications dealing with 3D reconstruction and multi-view video surveillance.

A Food Recognition System for Diabetic Patients based on an Optimized Bag of Features Model

Computer vision-based food recognition could be used to estimate a meal's carbohydrate content for diabetic patients. This study proposes a methodology for automatic food recognition, based on the Bag of Features (BoF) model. An extensive technical investigation was conducted for the identification and optimization of the best performing components involved in the BoF architecture, as well as the estimation of the corresponding parameters. For the design and evaluation of the prototype system, a visual dataset with nearly 5,000 food images was created and organized into 11 classes. The optimized system computes dense local features, using the scale-invariant feature transform on the HSV color space, builds a visual dictionary of 10,000 visual words by using the hierarchical k-means clustering and finally classifies the food images with a linear support vector machine classifier. The system achieved classification accuracy of the order of 78%, thus proving the feasibility of the proposed approach in a very challenging image dataset.

Food Recognition for Dietary Assessment Using Deep Convolutional Neural Networks

Diet management is a key factor for the prevention and treatment of diet-related chronic diseases. Computer vision systems aim to provide automated food intake assessment using meal images. We propose a method for the recognition of already segmented food items in meal images. The method uses a 6-layer deep convolutional neural network to classify food image patches. For each food item, overlapping patches are extracted and classified and the class with the majority of votes is assigned to it. Experiments on a manually annotated dataset with 573 food items justified the choice of the involved components and proved the effectiveness of the proposed system yielding an overall accuracy of 84.9%.

Higher-order regularization and morphological techniques for image segmentation

La segmentación de imágenes es un campo importante de la visión computacional y una de las áreas de investigación más activas, con aplicaciones en comprensión de imágenes, detección de objetos, reconocimiento facial, vigilancia de vídeo o procesamiento de imagen médica. La segmentación de imágenes es un problema difícil en general, pero especialmente en entornos científicos y biomédicos, donde las técnicas de adquisición imagen proporcionan imágenes ruidosas. Además, en muchos de estos casos se necesita una precisión casi perfecta. En esta tesis, revisamos y comparamos primero algunas de las técnicas ampliamente usadas para la segmentación de imágenes médicas. Estas técnicas usan clasificadores a nivel de pixel e introducen regularización sobre pares de píxeles que es normalmente insuficiente. Estudiamos las dificultades que presentan para capturar la información de alto nivel sobre los objetos a segmentar. Esta deficiencia da lugar a detecciones erróneas, bordes irregulares, configuraciones con topología errónea y formas inválidas. Para solucionar estos problemas, proponemos un nuevo método de regularización de alto nivel que aprende información topológica y de forma a partir de los datos de entrenamiento de una forma no paramétrica usando potenciales de orden superior. Los potenciales de orden superior se están popularizando en visión por computador, pero la representación exacta de un potencial de orden superior definido sobre muchas variables es computacionalmente inviable. Usamos una representación compacta de los potenciales basada en un conjunto finito de patrones aprendidos de los datos de entrenamiento que, a su vez, depende de las observaciones. Gracias a esta representación, los potenciales de orden superior pueden ser convertidos a potenciales de orden 2 con algunas variables auxiliares añadidas. Experimentos con imágenes reales y sintéticas confirman que nuestro modelo soluciona los errores de aproximaciones más débiles. Incluso con una regularización de alto nivel, una precisión exacta es inalcanzable, y se requeire de edición manual de los resultados de la segmentación automática. La edición manual es tediosa y pesada, y cualquier herramienta de ayuda es muy apreciada. Estas herramientas necesitan ser precisas, pero también lo suficientemente rápidas para ser usadas de forma interactiva. Los contornos activos son una buena solución: son buenos para detecciones precisas de fronteras y, en lugar de buscar una solución global, proporcionan un ajuste fino a resultados que ya existían previamente. Sin embargo, requieren una representación implícita que les permita trabajar con cambios topológicos del contorno, y esto da lugar a ecuaciones en derivadas parciales (EDP) que son costosas de resolver computacionalmente y pueden presentar problemas de estabilidad numérica. Presentamos una aproximación morfológica a la evolución de contornos basada en un nuevo operador morfológico de curvatura que es válido para superficies de cualquier dimensión. Aproximamos la solución numérica de la EDP de la evolución de contorno mediante la aplicación sucesiva de un conjunto de operadores morfológicos aplicados sobre una función de conjuntos de nivel. Estos operadores son muy rápidos, no sufren de problemas de estabilidad numérica y no degradan la función de los conjuntos de nivel, de modo que no hay necesidad de reinicializarlo. Además, su implementación es mucho más sencilla que la de las EDP, ya que no requieren usar sofisticados algoritmos numéricos. Desde un punto de vista teórico, profundizamos en las conexiones entre operadores morfológicos y diferenciales, e introducimos nuevos resultados en este área. Validamos nuestra aproximación proporcionando una implementación morfológica de los contornos geodésicos activos, los contornos activos sin bordes, y los turbopíxeles. En los experimentos realizados, las implementaciones morfológicas convergen a soluciones equivalentes a aquéllas logradas mediante soluciones numéricas tradicionales, pero con ganancias significativas en simplicidad, velocidad y estabilidad. ABSTRACT Image segmentation is an important field in computer vision and one of its most active research areas, with applications in image understanding, object detection, face recognition, video surveillance or medical image processing. Image segmentation is a challenging problem in general, but especially in the biological and medical image fields, where the imaging techniques usually produce cluttered and noisy images and near-perfect accuracy is required in many cases. In this thesis we first review and compare some standard techniques widely used for medical image segmentation. These techniques use pixel-wise classifiers and introduce weak pairwise regularization which is insufficient in many cases. We study their difficulties to capture high-level structural information about the objects to segment. This deficiency leads to many erroneous detections, ragged boundaries, incorrect topological configurations and wrong shapes. To deal with these problems, we propose a new regularization method that learns shape and topological information from training data in a nonparametric way using high-order potentials. High-order potentials are becoming increasingly popular in computer vision. However, the exact representation of a general higher order potential defined over many variables is computationally infeasible. We use a compact representation of the potentials based on a finite set of patterns learned fromtraining data that, in turn, depends on the observations. Thanks to this representation, high-order potentials can be converted into pairwise potentials with some added auxiliary variables and minimized with tree-reweighted message passing (TRW) and belief propagation (BP) techniques. Both synthetic and real experiments confirm that our model fixes the errors of weaker approaches. Even with high-level regularization, perfect accuracy is still unattainable, and human editing of the segmentation results is necessary. The manual edition is tedious and cumbersome, and tools that assist the user are greatly appreciated. These tools need to be precise, but also fast enough to be used in real-time. Active contours are a good solution: they are good for precise boundary detection and, instead of finding a global solution, they provide a fine tuning to previously existing results. However, they require an implicit representation to deal with topological changes of the contour, and this leads to PDEs that are computationally costly to solve and may present numerical stability issues. We present a morphological approach to contour evolution based on a new curvature morphological operator valid for surfaces of any dimension. We approximate the numerical solution of the contour evolution PDE by the successive application of a set of morphological operators defined on a binary level-set. These operators are very fast, do not suffer numerical stability issues, and do not degrade the level set function, so there is no need to reinitialize it. Moreover, their implementation is much easier than their PDE counterpart, since they do not require the use of sophisticated numerical algorithms. From a theoretical point of view, we delve into the connections between differential andmorphological operators, and introduce novel results in this area. We validate the approach providing amorphological implementation of the geodesic active contours, the active contours without borders, and turbopixels. In the experiments conducted, the morphological implementations converge to solutions equivalent to those achieved by traditional numerical solutions, but with significant gains in simplicity, speed, and stability.

Efficient hybrid monocular-stereo approach to on-board, video-based traffic sign detection and tracking

In this paper we propose an innovative method for the automatic detection and tracking of road traffic signs using an onboard stereo camera. It involves a combination of monocular and stereo analysis strategies to increase the reliability of the detections such that it can boost the performance of any traffic sign recognition scheme. Firstly, an adaptive color and appearance based detection is applied at single camera level to generate a set of traffic sign hypotheses. In turn, stereo information allows for sparse 3D reconstruction of potential traffic signs through a SURF-based matching strategy. Namely, the plane that best fits the cloud of 3D points traced back from feature matches is estimated using a RANSAC based approach to improve robustness to outliers. Temporal consistency of the 3D information is ensured through a Kalman-based tracking stage. This also allows for the generation of a predicted 3D traffic sign model, which is in turn used to enhance the previously mentioned color-based detector through a feedback loop, thus improving detection accuracy. The proposed solution has been tested with real sequences under several illumination conditions and in both urban areas and highways, achieving very high detection rates in challenging environments, including rapid motion and significant perspective distortion

Image recognition: Visual grouping, recognition, and learning

Vision extracts useful information from images. Reconstructing the three-dimensional structure of our environment and recognizing the objects that populate it are among the most important functions of our visual system. Computer vision researchers study the computational principles of vision and aim at designing algorithms that reproduce these functions. Vision is difficult: the same scene may give rise to very different images depending on illumination and viewpoint. Typically, an astronomical number of hypotheses exist that in principle have to be analyzed to infer a correct scene description. Moreover, image information might be extracted at different levels of spatial and logical resolution dependent on the image processing task. Knowledge of the world allows the visual system to limit the amount of ambiguity and to greatly simplify visual computations. We discuss how simple properties of the world are captured by the Gestalt rules of grouping, how the visual system may learn and organize models of objects for recognition, and how one may control the complexity of the description that the visual system computes.

Face encoding and recognition in the human brain.

A dissociation between human neural systems that participate in the encoding and later recognition of new memories for faces was demonstrated by measuring memory task-related changes in regional cerebral blood flow with positron emission tomography. There was almost no overlap between the brain structures associated with these memory functions. A region in the right hippocampus and adjacent cortex was activated during memory encoding but not during recognition. The most striking finding in neocortex was the lateralization of prefrontal participation. Encoding activated left prefrontal cortex, whereas recognition activated right prefrontal cortex. These results indicate that the hippocampus and adjacent cortex participate in memory function primarily at the time of new memory encoding. Moreover, face recognition is not mediated simply by recapitulation of operations performed at the time of encoding but, rather, involves anatomically dissociable operations.

The Illinois pattern recognition computer (Illiac III) /

"Presented at the 18th Annual ACM National Conference, Denver, Colorado August 27, 1963"

Low-cost real-time gesture recognition

A major impediment to developing real-time computer vision systems has been the computational power and level of skill required to process video streams in real-time. This has meant that many researchers have either analysed video streams off-line or used expensive dedicated hardware acceleration techniques. Recent software and hardware developments have greatly eased the development burden of realtime image analysis leading to the development of portable systems using cheap PC hardware and software exploiting the Multimedia Extension (MMX) instruction set of the Intel Pentium chip. This paper describes the implementation of a computationally efficient computer vision system for recognizing hand gestures using efficient coding and MMX-acceleration to achieve real-time performance on low cost hardware.

Probabilistic visual recognition of artificial landmarks for simultaneous localization and mapping

Probabilistic robotics most often applied to the problem of simultaneous localisation and mapping (SLAM), requires measures of uncertainty to accompany observations of the environment. This paper describes how uncertainty can be characterised for a vision system that locates coloured landmarks in a typical laboratory environment. The paper describes a model of the uncertainty in segmentation, the internal cameral model and the mounting of the camera on the robot. It explains the implementation of the system on a laboratory robot, and provides experimental results that show the coherence of the uncertainty model.