Catálogo de publicaciones - libros

3-D reconstruction from images sequences has been the center topic of computer vision. Real-time applications call for causal processing of stereo sequences, as they are acquired, covering different regions of the scene. The first step is to compute the current stereo disparity, and recursive map building often requires fusing with the previous estimate. In this paper, the epiflow framework [1], originally proposed for establishing matches among stereo feature pairs is generalized to devise an iterative causal algorithm for stereo disparity map fusion. In the context of disparity fusion, quadruplet correspondence of the epiflow tracking algorithm becomes reminiscent of the “closest point” of the 3-D ICP algorithm. Unlike ICP, the 2-D epiflow framework permits incorporating both photometric and geometrical constraints, estimation of the stereo rig motion as supplementary information, as well as identifying local inconsistencies between the two disparity maps. Experiments with real data validate the proposed approach, and improved converge compared to the ICP algorithm.

Liver volumetry is a required step for the planning of liver surgery and resection. It is generally based on Computerized tomography images, and segmentation of the liver is the most important step of the process. We propose an automatic segmentation algorithm based on a geometric level set method which provides an accurate segmentation of the liver, and requires no a priori information. We show results on different datasets, with and without a contrast agent. The segmentation is compared to manual delineation by a radiologist with good results.

This paper presents a method to decompose a field of surface normals (needle-map). A diffusion process is used to model the flow of height information induced by a field of surface normals. The diffusion kernel can be decomposed into eigenmodes, each corresponding to approximately independent modes of variation of the flow. The surface normals can then be diffused using a modified kernel with the same eigenmodes but different coefficients. When used as part of a surface integration process, this procedure allows choosing the trade-off between local and global influence of each eigenmode in the modified field of surface normals. This graph-spectral method is illustrated with surface normals extracted from a face. Experiments are carried with local affinity functions that convey both the intrinsic and extrinsic geometry of the surface, and an information-theoretic definition of affinity.

This paper deals with the evaluation of the performance of a set of classifiers on multispectral imagery with low dimensionality and low spatial and spectral resolutions. The original Landsat TM images and other 4 transformed sets are classified by 5 supervised and 2 unsupervised methods. The results for 7 land cover classes are compared and the performances of the methods for each set of input data are discussed.

An Audio fingerprint is a small digest of an audio file computed from its main perceptual properties. Like human fingerprints, audio fingerprints allow to identify an audio file among a set of candidates without retrieving any other characteristics. We propose in this paper a fingerprint extraction algorithm based on a new audio segmentation method. A new scoring function based on q-grams is used to determine if an input signal is a derivated version of a fingerprint stored in the database. A rule based on this scoring function allows to either recover the original input file or to decide that no fingerprint belonging to the database corresponds to the signal. The proposed method is robust against compression and time shifting alterations of audio files.

This paper presents a methodology for the automated detection of impact craters on images of planetary surfaces. This modular approach includes a phase of candidate selection, followed by template matching and finally the analysis of a probability volume that allows for the identification of craters on the image. It is applied to a set of images of the surface of the planet Mars, with results that are very promising, in face of future improvements in the methodology.

In human motion analysis, the joint estimation of appearance, body pose and location parameters is not always tractable due to its huge computational cost. In this paper, we propose a Rao-Blackwellized Particle Filter for addressing the problem of human pose estimation and tracking. The advantage of the proposed approach is that Rao-Blackwellization allows the state variables to be splitted into two sets, being one of them analytically calculated from the posterior probability of the remaining ones. This procedure reduces the dimensionality of the Particle Filter, thus requiring fewer particles to achieve a similar tracking performance. In this manner, location and size over the image are obtained stochastically using colour and motion clues, whereas body pose is solved analytically applying learned human Point Distribution Models.

In this paper, two systems of physiological parameters that describe the human functional state are analysed. Some physiological features are measured on sportsmen and ischemic heart-diseased men by changing the physical load, and the system of parameters is developed. One parameter system is based on the so-called fractal dimensions, and the other one is based on the parameters of curves that approximate the change of feature values in time. The application of classification and visualization methods and neural networks allowed us to optimise the number of parameters. The results may be generalised to the large number of problems in biomedicine and bioinformatics.

Color has plenty of discriminative information that can be used to improve the performance of face recognition algorithms, although it is difficult to use it because of its high variability. In this paper we investigate the use of the quaternion representation of a color image for face recognition. We also propose a new representation for color images based on complex numbers. These two color representation methods are compared with the traditional grayscale and RGB representations using an eigenfaces based algorithm for identity verification. The experimental results show that the proposed method gives a very significant improvement when compared to using only the illuminance information.

Zoologists studying elephant populations in wild environments need to recognize different individuals from photos taken in different periods of time. Individuals can be distinguished by the shape of the nicks on their ears. Nevertheless, shape comparison is not trivial due to a highly cluttered background. We propose a method for partially, non-connected curve matching able to compare photos of elephant ears.