Catálogo de publicaciones - libros

In this paper, we propose a novel method for estimating major soccer scenes from cameraworks and players trajectories based on probabilistic inference, and annotating scene indexes to broadcast soccer videos automatically. In our method, we define relations between cameraworks and scenes, and between players trajectories and scenes by conditional probabilities. Moreover defining temporal relations of scenes by transition probabilities, we represent those relations as dynamic bayesian networks (DBNs). And those probabilities are evaluated by learning parameters of the networks. After extracting the cameraworks and the players trajectories, we compute the posterior probability distribution of scenes, and give the computed results to the soccer video as the scene index. Finally, we discuss the extendibility of the proposal indexing technique in the case of adding ball trajectories and audios.

Self-organized maps have been proposed as a model for the formation of sensory maps in the cerebral cortex. The role of inputs is critical in this process of self-organization. This paper presents a systematic approach to analyzing the relationship between the input ensemble and the quality of self-organization achieved.

We present a method for generating an input stimulus set consisting of images of curved lines. The advantage of this approach is that it allows the user the ability to precisely control the statistics of the input stimuli to visual processing algorithms. Since there is considerable scientific interest in the processing of information in the human visual stream, we specifically address the problem of self-organization of cortical visual areas V1 and V2.

We show that the statistics of the curves generated with our algorithm match the statistics of natural images. We develop a measure of self-organization based on the oriented energy contained in the afferent weights to each cortical unit in the map. We show that as the curvature of the generated lines increases, this measure of self-organization decreases. Furthermore, self-organization using curved lines as stimuli is achieved much more rapidly, as the curve images do not contain as much higher order structure as natural images do.

The usual goal of modeling natural and artificial perception involves determining how a system can extract the object that it perceives from an image which is noisy. The “inverse” of this problem is one of modeling how even a image can be perceived to be blurred in certain contexts. We propose a chaotic model of Pattern Recognition (PR) for the theory of “blurring”. The paper, which is an extension to a Companion paper [3] demonstrates how one can model blurring from the view point of a chaotic PR system. Unlike the Companion paper in which the chaotic PR system extracts the pattern from the input, this paper shows that the perception can be “blurred” if the dynamics of the chaotic system are modified. We thus propose a formal model, the Mb-AdNN, and present a rigorous analysis using the Routh-Hurwitz criterion and Lyapunov exponents. We also demonstrate, experimentally, the validity of our model by using a numeral dataset.

We present several methods for per-region land-cover classification based on distances on probability distributions and whole-region probabilities. We present results on using this method for locating forest areas in high-resolution aerial images with very high reliability, achieving more than 95% accuracy, using raw radiometric channels as well as derived color and texture features. Region boundaries are obtained from a multi-scale hierarchical segmentation or from a registration of cadastral maps.

We address the problem of incorporating transformation invariance in kernels for pattern analysis with kernel methods. We introduce a new class of kernels by so called Haar-integration over transformations. This results in kernel functions, which are positive definite, have adjustable invariance, can capture simultaneously various continuous or discrete transformations and are applicable in various kernel methods. We demonstrate these properties on toy examples and experimentally investigate the real-world applicability on an image recognition task with support vector machines. For certain transformations remarkable complexity reduction is demonstrated. The kernels hereby achieve state-of-the-art results, while omitting drawbacks of existing methods.

This paper presents an approach of visual shape recognition based on exemplars of attributed keypoints. Training is performed by storing exemplars of keypoints detected in labeled training images. Recognition is made by keypoint matching and voting according to the labels for the matched keypoints. The matching is insensitive to rotations, limited scalings and small deformations. The recognition is robust to noise, background clutter and partial occlusion. Recognition is possible from few training images and improve with the number of training images.

This paper deals, for the first time, with an analysis of localization capabilities of weakly supervised categorization systems. Most existing categorization approaches have been tested on databases, which (a) either show the object(s) of interest in a very prominent way so that their localization can hardly be judged from these experiments, or (b) at least the learning procedure was done with supervision, which forces the system to learn only object relevant data. These approaches cannot be directly compared to a nearly unsupervised method. The main contribution of our paper thus is twofold: First, we have set up a new database which is sufficiently complex, balanced with respect to background, and includes localization ground truth. Second, we show, how our successful approach for generic object recognition [14] can be extended to perform localization, too.To analyze its localization potential, we develop localization measures which focus on approaches based on Boosting [5]. Our experiments show that localization depends on the object category, as well as on the type of the local descriptor.

Clustering algorithms are intensively used in the image analysis field in compression, segmentation, recognition and other tasks. In this work we present a new approach in clustering vector datasets by finding a good order in the set, and then applying an optimal segmentation algorithm. The algorithm heuristically prolongs the optimal scalar quantization technique to vector space. The data set is sequenced using one-dimensional projection spaces. We show that the principal axis is too rigid to preserve the adjacency of the points. We present a way to refine the order using the minimum weight Hamiltonian path in the data graph. Next we propose to use the principal curve to better model the non-linearity of the data and find a good sequence in the data. The experimental results show that the principal curve based clustering method can be successfully used in cluster analysis.

In this paper, two block-based texture methods are proposed for content-based image retrieval (CBIR). The approaches use the Local Binary Pattern (LBP) texture feature as the source of image description. The first method divides the query and database images into equally sized blocks from which LBP histograms are extracted. Then the block histograms are compared using a relative dissimilarity measure based on the Minkowski distances. The second approach uses the image division on database images and calculates a single feature histogram for the query. It sums up the database histograms according to the size of the query image and finds the best match by exploiting a sliding search window. The first method is evaluated against color correlogram and edge histogram based algorithms. The second, user interaction dependent approach is used to provide example queries. The experiments show the clear superiority of the new algorithms against their competitors.

The shapes occurring in the images are essential features in image classification and retrieval. Due to their compactness and classification accuracy, Fourier-based shape descriptors are popular boundary-based methods for shape description. However, in the case of short boundary functions, the frequency resolution of the Fourier spectrum is low, which yields to inadequate shape description. Therefore, we have applied zero-padding method for the short boundary functions to improve their Fourier-based shape description. In this paper, we show that using this method the Fourier-based shape classification can be significantly improved.