Catálogo de publicaciones - libros

This paper investigates the use of local descriptors that are based on integral invariants for the purpose of object recognition in cluttered scenes. Integral invariants capture the local structure of the neighborhood around the points where they are computed. This makes them very well suited for constructing highly-discriminative local descriptors. The features are by definition invariant to Euclidean motion. We show how to extend the local features to be scale invariant. Regarding the robustness to intensity changes, two types of kernels used for extracting the feature vectors are investigated. The effect of the feature vector dimensionality and the performance in the presence of noise are also examined. Promising results are obtained using a dataset that contains instances of objects that are viewed in difficult situations that include clutter and occlusion.

The goal of this paper is to study the set of features that is suitable for describing animals in images, and for being able to categorize them in natural scenes. We propose multi-scale features based on Gaussian derivatives functions, that show interesting invariance properties. In order to build an efficient system, we will use classifiers based on the JointBoosting methodology, which will be compared with the well-known one-vs-all approach by using Support Vector Machines. Thirty five categories, containing animals, are selected from the challenging Caltech 101 object categories database to carry out the study.

This paper presents a method for object categorization. This problem is difficult and can be solved by combining different information sources such as shape or appearance. In this paper, we aim at performing object recognition by mixing kernels obtained from different cues. Our method is based on two complementary descriptions of an object. First, we describe its shape thanks to labeled graphs. This graph is obtained from morphological skeleton, extracted from the binary mask of the object image. The second description uses histograms of oriented gradients which aim at capturing objects appearance. The histogram descriptor is obtained by computing local histograms over the complete image of the object. These two descriptions are combined using a kernel product. Our approach has been validated on the ETH80 database which is composed of 3280 images gathered in 8 classes. The results we achieved show that this method can be very efficient.

In this work, four alternative algorithms for Classification/ Discrimination are described which we have developed recently. These are believed to have several advantages over the previous ones via the introduction of a new concept, Centers of Masses for classes, and new cost functions which enforce clustering of different classes more explicitly compared to the previous approaches.

Dynamics of training the group of single layer perceptrons aimed to solve multi-class pattern recognition problem is studied. It is shown that in special training of the perceptrons, one may obtain a pool of different classification algorithms. Means to improve training speed and reduce generalization error are studied. Training dynamics is illustrated by solving artificial multi-class pattern recognition task and important real world problem: detection of ten types of yeast infections from 1500 spectral features.

In this paper a new formulation of probabilistic relaxation labeling is developed using the theory of diffusion processes on graphs. Our idea is to formulate relaxation labelling as a diffusion process on the vector of object-label probabilities. According to this picture, the label probabilities are given by the state-vector of a continuous time random walk on a support graph. The state-vector is the solution of the heat equation on the support-graph. The nodes of the support graph are the Cartesian product of the object-set and label-set of the relaxation process. The compatibility functions are combined in the weight matrix of the support graph. The solution of the heat-equation is found by exponentiating the eigensystem of the Laplacian matrix for the weighted support graph with time. We demonstrate the new relaxation process on a toy labeling example which has been studied extensively in the early literature, and a feature correspondence matching problem abstracted in terms of relational graphs. The experiments show encouraging labeling and matching results.

This paper presents a framework for gesture recognition by modeling a system based on Dynamic Bayesian Networks (DBNs) from a Marionette point of view. To incorporate human qualities like anticipation and empathy inside the perception system of a social robot remains, so far an open issue. It is our goal to search for ways of implementation and test the feasibility. Towards this end we started the development of the guide robot ’Nicole’ equipped with a monocular camera and an inertial sensor to observe its environment. The context of interaction is a person performing gestures and ’Nicole’ reacting by means of audio output and motion. In this paper we present a solution to the gesture recognition task based on Dynamic Bayesian Network (DBN). We show that using a DBN is a human-like concept of recognizing gestures that encompass the quality of anticipation through the concept of prediction and update. A novel approach is used by incorporating a marionette model in the DBN as a trade-off between simple constant acceleration models and complex articulated models.

As pattern recognition methods, subspace methods have attracted much attention in the fields of face, object and video-based recognition in recent years. In subspace methods, each instance is characterized by a subspace that is spanned by a set of vectors. Thus, the distance between instances reduces to the distance between subspaces. Herein, the subspace distance designing problem is considered mathematically. Any distance designed according the method presented here can be embedded into associated recognition algorithms. The main contributions in this paper include: – Solving the open problem proposed by Wang, Wang and Feng (2005), that is, we proved that their dissimilarity is a distance; – Presenting a general framework of subspace construction, concretely speaking, we pointed out a view that subspace distance also could be regarded as the classical distance in vector space; – Proposing two types of kernel subspace distances; – Comparing some known subspace (dis)similarities mathematically.

Automatic classification of objects based on their visual appearance is often performed based on clustering algorithms, which can be based on soft computing techniques. One of the most used methods is fuzzy clustering. However, this method can converge to local minima. This problem has been addressed very recently by applying ant colony optimization to tackle this problem. This paper proposed the use of this fuzzy-ant clustering approach to derive fuzzy models. These models are used to classify marbles based on their visual appearance; color and vein classification is performed. The proposed fuzzy modeling approach is compared to other soft computing classification algorithms, namely: fuzzy, neural, simulated annealing, genetic and combinations of these approaches. Fuzzy-ant models presented higher classification rates than the other soft computing techniques.

In the field of cluster analysis, most of existing algorithms assume that each feature of the samples plays a uniform contribution for cluster analysis and they are mainly developed for the uniform distribution of sample with numerical or categorical data, which cannot effectively process the non-uniformly distribution with mixed data sets in data mining. For this purpose, we propose a new general Weighted Fuzzy Clustering Algorithm to deal with the mixed data including different sample distributions and different features, in which the idea of the probability density of samples is used to assign the weights to each sample and the ReliefF algorithms is applied to give the weights to each feature. By weighting the samples and their features, the fuzzy c-means, fuzzy c -modes, fuzzy c -prototype and sample-weighted algorithms can be unified into a general framework. The experimental results with various test data sets illustrate the effectiveness of the proposed clustering algorithm.