Catálogo de publicaciones - libros

In time-series analysis it is often assumed that observed data can be modelled as being derived from a number of regimes of dynamics, as e.g. in a Switching Kalman Filter (SKF) [8,2]. However, it may not be possible to model all of the regimes, and in this case it can be useful to represent explicitly a ‘novel’ regime. We apply this idea to the Factorial Switching Kalman Filter (FSKF) by introducing an extra factor (the ‘X-factor’) to account for the unmodelled variation. We apply our method to physiological monitoring data from premature infants receiving intensive care, and demonstrate that the model is effective in detecting abnormal sequences of observations that are not modelled by the known regimes.

Analysis, interpretation and manipulation of complex visual data is something which humans do quite easily. We can easily recognize different instances of objects (like faces, bodies, flowers, etc.), in spite of the huge variations in their appearances.We have no problem determining that two actions are the same, even though they are performed by different people wearing different clothes against different backgrounds.Moreover, we humans can predict and make inferences about very complex static and dynamic visual information which we have never seen. We can infer about the appearance of unfamiliar places, objects and actions, as well as detect saliency and abnormalities in such data. Such complex visual tasks require sophisticated notions of visual similarity and dissimilarity.

When a world-observing camera moves through a scene capturing images continuously, it is possible to analyse the images to estimate its , even if nothing is known in advance about the contents of the scene around it. The key to solving this apparently chicken-and-egg problem is to detect and repeatedly measure a number of salient ‘features’ in the environment as the camera moves. Under the usual assumption that most of these are rigidly related in the world, the many geometric constraints on relative camera/feature locations provided by image measurements allow one to solve simultaneously for both the camera motion and the 3D world positions of the features. While global optimisation algorithms are able to achieve the most accurate solutions to this problem, the consistent theme of my research has been to develop ‘Simultaneous Localisation and Mapping’ (SLAM) algorithms using probabilistic filtering which permit sequential, hard operation.

One of the major difficulties of handwriting recognition is the variability among symbols because of the different writer styles. In this paper we introduce the boosting of blurred shape models with error correction, which is a robust approach for describing and recognizing handwritten symbols tolerant to this variability. A symbol is described by a probability density function of blurred shape model that encodes the probability of pixel densities of image regions. Then, to learn the most distinctive features among symbol classes, boosting techniques are used to maximize the separability among the blurred shape models. Finally, the set of binary boosting classifiers is embedded in the framework of Error Correcting Output Codes (ECOC). Our approach has been evaluated in two benchmarking scenarios consisting of handwritten symbols. Compared with state-of-the-art descriptors, our method shows higher tolerance to the irregular deformations induced by handwritten strokes.

This paper presents a new Bayesian approach to hyperspectral image segmentation that boosts the performance of the discriminative classifiers. This is achieved by combining class densities based on discriminative classifiers with a Multi-Level Logistic Markov-Gibs prior. This density favors neighbouring labels of the same class. The adopted discriminative classifier is the Fast Sparse Multinomial Regression. The discrete optimization problem one is led to is solved efficiently via graph cut tools. The effectiveness of the proposed method is evaluated, with simulated and real AVIRIS images, in two directions: 1) to improve the classification performance and 2) to decrease the size of the training sets.

Different methods have been proposed in order to deal with the huge amount of information that hyperspectral applications involve. This paper presents a comparison of some of the methods proposed for band selection. A relevant and recent set of methods have been selected that cover the main tendencies in this field. Moreover, a variant of an existing method is also introduced in this work. The comparison criterion used is based on pixel classification tasks.

The aim in this paper is to show how to discriminate gender using a parameterized representation of fields of facial surface normals (needle-maps). We make use of principle geodesic analysis (PGA) to parameterize the facial needle-maps. Using feature selection, we determine the selected feature set which gives the best result in distinguishing gender. Using the EM algorithm we distinguish gender by fitting a two component mixture model to the vectors of selected features. Results on real-world data reveal that the method gives accurate gender discrimination results.

This paper proposes an incremental method for feature selection, aimed at identifying attributes in a dataset that allow to buid classifiers at low computational cost. The basis of the approach is the minimal-redundancy-maximal-relevance (mRMR) framework, which attempts to select features relevant for a given classification task, avoiding redundancy among them. Relevance and redundancy have been popularly defined in terms of information theory concepts. In this paper a modification of the mRMR framework is proposed, based on a more proper quantification of the redundancy among features. Experimental work on discrete–valued datasets shows that classifiers built using features selected by the proposed method are more accurate than the ones obtained using original mRMR features.

One main process in Computer Vision is image segmentation as a tool to other visual tasks. Although there are many approaches to grey scale image segmentation, nowadays most of the digital images are colour images. This paper introduces a new method for colour image segmentation. We focus our work on a topological study of colour distribution, e.g., image histogram. We argue that this point of view bring us the possibility to find dominant colours by preserving the spatial coherence of the histogram. To achieve it, we find and extract ridges of the colour distribution and assign a unique colour at every ridge as a representative colour of an interest region. This method seems to be not affected by shadows in a wide range of tested images.

We address the problem of non-linearity in 2D Shape modelling of a particular articulated object: the human body. This issue is partially resolved by applying a different Point Distribution Model (PDM) depending on the viewpoint. The remaining non-linearity is solved by using Gaussian Mixture Models (GMM). A dynamic-based clustering is proposed and carried out in the Pose Eigenspace. A fundamental question when clustering is to determine the optimal number of clusters. From our point of view, the main aspect to be evaluated is the mean gaussianity. This partitioning is then used to fit a GMM to each one of the view-based PDM, derived from a database of Silhouettes and Skeletons. Dynamic correspondences are then obtained between gaussian models of the 4 mixtures. Finally, we compare this approach with other two methods we previously developed to cope with non-linearity: Nearest Neighbor (NN) Classifier and Independent Component Analysis (ICA).