Catálogo de publicaciones - libros

We have developed an efficient algorithm to compute an Euclidean reconstruction from only two wide-baseline color images captured with a hand-held digital camera. The classical reconstruction scheme has been improved to boost the number of matches by a hierarchical epipolar constraint during an iterative process and an ultimate step of dense matching based on affine transformation. At the output, between three to four thousands points are reconstructed in 2 minutes on 1024x768 images. The stability of the algorithm has been evaluated by some repetitive tests and the quality of the reconstruction is assessed according to a metric ground truth provided by an industrial 3D scanner. The averaged error on 3D points is around 3.5% reported to the model depth. Such a precision makes this technique suitable for wound volumetric assessment in clinical environments using a hand held digital camera.

Today most mobile telephones come equipped with a camera. This gives rise to interesting new possibilities for applications of computer vision, such as building recognition software running locally on the mobile phone. Algorithms for building recognition need to be robust under noise, occlusion, varying lighting conditions and different points of view. We present such an algorithm using local invariant regions which allows for mobile building recognition despite the limited processing power and storage capacity of mobile phones. This algorithm was shown to obtain state of the art performance on the Zürich Building Database (91% accuracy). An implementation on a mobile phone (Sony Ericsson K700i) is presented that obtains good performance (80% accuracy) on a dataset using real-world query images taken under varying, suboptimal conditions. Our algorithm runs in the order of several seconds while requiring only around 10 KB of memory to represent a single building within the local database.

This paper presents a novel on-line learning procedure to be used in biologically realistic networks of integrate-and-fire neurons. The on-line adaptation is based on synaptic plasticity and changes in the network structure. Event driven computation optimizes processing speed in order to simulate networks with large number of neurons. The learning method is demonstrated on a visual recognition task and can be expanded to other data types. Preliminary experiments on face image data show the same performance as the optimized off-line method and promising generalization properties.

The vision-based scene understanding technique that infers scene-interpreting contexts from real-world vision data has to not only deal with various uncertain environments but also reflect user’s requests. Especially, learnability is a hot issue for the system. In this paper, we adopt a probabilistic approach to overcome the uncertainty, and propose an interactive learning method using combination of Bayesian network and logic network to reflect user’s requirements in real-time. The logic network works for supporting logical inference of Bayesian network. In the result of some learning experiments using interactive data, we have confirmed that the proposed interactive learning method is useful for scene context reasoning.

This paper presents a Road Detection and Classification algorithm for Driver Assistance Systems (DAS), which tracks several road lanes and identifies the type of lane boundaries. The algorithm uses an edge filter to extract the longitudinal road markings to which a straight lane model is fitted. Next, the type of right and left lane boundaries (, or line) is identified using a Fourier analysis. Adjacent lanes are searched when broken or merge lines are detected. Although the knowledge of the line type is essential for a robust DAS, it has been seldom considered in previous works. This knowledge helps to guide the search for other lanes, and it is the basis to identify the type of road (, or ), as well as to tell the difference between allowed and forbidden maneuvers, such as crossing a continuous line.

A fast method for the recognition and classification of informational traffic signs is presented in this paper. The aim is to provide an efficient framework which could be easily used in inventory and guidance systems. The process consists of several steps which include image segmentation, sign detection and reorientation, and finally traffic sign recognition. In a first stage, a static HSI colour segmentation is performed so that possible traffic signs can be easily isolated from the rest of the scene; secondly, shape classification is carried out so as to detect square blobs from the segmented image; next, each object is reoriented through the use of a homography transformation matrix and its potential axial deformation is corrected. Finally a recursive adaptive segmentation and a SVM-based recognition framework allow us to extract each possible pictogram, icon or symbol and classify the type of the traffic sign via a voting-scheme.

For high-dimensional classification tasks such as face recognition, the number of samples is smaller than the dimensionality of the samples. In such cases, a problem encountered in Linear Discriminant Analysis-based (LDA) methods for dimension reduction is what is known as the Small Sample Size (SSS) problem. Recently, a number of approaches that attempt to solve the SSS problem have been proposed in the literature. In this paper, a different way of solving the SSS problem compared to these is proposed. It is one that employs a dissimilarity representation method where an object is represented based on the dissimilarity measures among representatives extracted from training samples instead of from the feature vector itself. Thus, by appropriately selecting representatives and by defining the dissimilarity measure, it is possible to reduce the dimensionality and achieve a better classification performance in terms of both speed and accuracy. Apart from utilizing the dissimilarity representation, in this paper simultaneously employing a fusion technique is also proposed in order to increase the classification accuracy. The rationale for this is explained in the paper. The proposed scheme is completely different from the conventional ones in terms of the computation of the transformation matrix, as well as in controlling the number of dimensions. The present experimental results, which to the best of the authors’ knowledge, are the first such reported results, demonstrate that the proposed mechanism achieves nearly identical efficiency results in terms of the classification accuracy compared with the conventional LDA-extension approaches for well-known face databases involving AT&T and Yale databases.

The nearest neighbor (NN) classifier is well suited for generic object recognition. However, it requires storing the complete training data, and classification time is linear in the amount of data. There are several approaches to improve runtime and/or memory requirements of nearest neighbor methods: Thinning methods select and store only part of the training data for the classifier. Efficient query structures reduce query times. In this paper, we present an experimental comparison and analysis of such methods using the ETH-80 database. We evaluate the following algorithms. Thinning: condensed nearest neighbor, reduced nearest neighbor, Baram’s algorithm, the Baram-RNN hybrid algorithm, Gabriel and GSASH thinning. Query structures: kd-tree and approximate nearest neighbor. For the first four thinning algorithms, we also present an extension to -NN which allows tuning the trade-off between data reduction and classifier degradation. The experiments show that most of the above methods are well suited for generic object recognition.

Independent Component Analysis (ICA) has shown success in blind source separation. Its applications to remotely sensed images have been investigated recently. In this approach, a Linear Spectral Mixture (LSM) model is used to characterize spectral data. This model and the associated linear unmixing algorithms are based on the assumption that the spectrum for a given pixel in an image is a linear combination of the end-member spectra. The assumption that the abundances are mutually statistically independent random sources requires the separating matrix to be unitary. This paper considers a new approach, the Non Orthogonal Component Analysis (NOCA), which enables to relax this assumption. The experimental results demonstrate that the proposed NOCA provides a more effective technique for anomaly detection in hyperspectral imagery than the ICA approach. In particular, we highlight the fact that the difference between the performances of the two approaches increases when the number of bands decreases.

Video classification provides an efficient way to manage and utilize the video data. Existing works on this topic fall into this category: enlarging the feature set until the classification is reliable enough. However, some features may be redundant or irrelevant. In this paper, we address the problem of choosing efficient feature set in video genre classification to achieve acceptable classification results but relieve computation burden significantly. A rough set approach is proposed. In comparison with existing works and the decision tree method, experimental results verify the efficiency of the proposed approach.