Catálogo de publicaciones - libros

In this paper we present two new sensor systems for 3D data acquisition. Both systems are primarily designed for face scanning applications, but their performance characteristics are useful for technical applications as well. Their operating principle is based on structured light, more precisely on color coded light, implemented in the invisible near infrared spectrum. The resulting 3D data acquisition works in real-time, shows high robustness against ambient light and provides accurate high resolution depth maps.

Thus, most of the limiting factors known from other face scanning systems are eliminated, such as visible annoying and blinding light patterns, motion constraints and highly restricted application scenarios due to ambient light constraints.

This paper develops a new statistical model that can be used to recover the 3D face shape from a single image taken under arbitrary lighting conditions. The proposed statistical model combines shape and image intensity information. The latter is represented by a spherical harmonics (SH) space. Given a training data set of aligned height maps of faces and their corresponding albedos, The input image is projected into the space spanned by the SH basis images of each member in the data set. The combined statistical model is obtained by performing the PCA on both the SH projections and the height maps. Finally, the face is reconstructed by fitting the model to the input intensity image. In addition to the shape recovery, the proposed model can be used to recover the face albedo and to estimate the light direction. Experiments have been conducted to evaluate the performance of the proposed approach.

We present an intuitive, fast and accurate interactive segmentation method for extracting and visualizing a large range of objects from 3D medical images. Our method combines a general deformable subdivision-surface model with a novel sketch-line user initialization process. The model is simply and precisely initialized with a few quick sketch lines drawn across the width of the target object on several key slices of the volume image. The smooth surface constructed using these lines is extremely close to the shape of the object boundary, making the model’s task of snapping to this boundary much simpler and hence more likely to succeed in noisy images with minimal user editing. Our subdivision surface model provides a foundation for precise user steering/editing capabilities and simple, intuitive user interactions are seamlessly integrated with advanced visualization capabilities. We use our model to segment objects from several 3D medical images to demonstrate its effectiveness.

Resource allocation in ad hoc communication networks is a field of high complexity because of both ) the distributed nature of the interactions between the nodes, and ) the large set of control variables for even the most primitive networks. Visual representation of this information across physical space and across layers of the network can greatly benefit the understanding of efficient allocation policies in these complex systems. Yet, very few software packages have been developed to address specifically this task, especially for large scale networks. We develop such a software system, and demonstrate some of its capabilities. The system illustrates that multi-layered visualization of the network state can be an effective tool at making network design decisions even in the face of uncertainty and for a large number of network interactions.

Intrusion detection systems gather large quantities of host and network information in an attempt to detect and respond to attacks against an organization. The widely varying nature of attacks makes humans essential for analysis, but the sheer volume of data can quickly overwhelm even experienced analysts. Existing approaches utilize visualization to provide rapidly comprehensible representations of the data, but fail to scale to real-world environments due to unrealistic data handling and lack of response facilities. This paper introduces a new tool for security event monitoring, analysis, and response called Savors. Savors provides suitable scalability by utilizing three additional areas of computing. High-end computing brings large amounts of on-demand processing to bear on the problem. Auralization allows both monitoring and analysis to be performed in parallel. Finally, grid computing provides the basis for remote data access and response capabilities with seamless and secure access to organization resources.

In this paper, we present a method for analyzing the complexity of information visualization. We have identified a number of factors that influence the efficiency of visual information read-off and integration. Here the complexity is measured in terms of visual integration, number of separable dimensions for each visual unit, the complexity of interpreting the visual attributes, and the efficiency of visual search. These measures are derived from well established psychological theories. Together they indicate the amount of cognitive load for comprehending a visualization design. This method is particularly useful for developers to quickly evaluate multiple design choices. Finally, we demonstrate our method through a complexity analysis on a computer security visualization program.

This paper describes a GPU framework for the real-time visualization of natural textured terrains, as well as the steps that are needed to populate them on-the-fly with tens of thousands of plant and/or mineral objects. Our main contribution is a robust modular architecture developed for the G80 and later GPUs, that performs texture/seed selection and rendering. It does not deal with algorithms that procedurally model or render either terrain or specific natural objects, but uses them for demonstration purposes. It can be used to calculate and display realistic landscapes and ecosystems with minimal pre-stored data, CPU load and popping artefacts. It works in tandem with a pre-classification of available aerial images, and makes it possible to fine-tune the properties of objects added to the scene.

The ray-casting of implicit surfaces on GPU has been explored in the last few years. However, until recently, they were restricted to second degree (quadrics). We present an iterative solution to ray cast cubics and quartics on GPU. Our solution targets efficient implementation, obtaining interactive rendering for thousands of surfaces per frame. We have given special attention to torus rendering since it is a useful shape for multiple CAD models. We have tested four different iterative methods, including a novel one, comparing them with classical tessellation solution.

In this work, we present a novel, fast clustering scheme for codebook generation from local features for object class recognition. It relies on a sequential data analysis and creates compact clusters with low variance. We compare our algorithm to other commonly used algorithms with respect to cluster statistics and classification performance. It turns out that our algorithm is the fastest for codebook generation, without loss in classification performance, when using the right matching scheme. In this context, we propose a well suited matching scheme for assigning data entries to cluster centers based on the sigmoid function.

The iris is currently accepted as one of the most accurate traits for biometric purposes. However, for the sake of accuracy, iris recognition systems rely on good quality images and significantly deteriorate their results when images contain large noisy regions, either due to iris obstructions (eyelids or eyelashes) or reflections (specular or lighting). In this paper we propose an entropy-based iris coding strategy that constructs an unidimensional signal from overlapped angular patches of normalized iris images. Further, in the comparison between biometric signatures we exclusively take into account signatures’ segments of varying dimension. The hope is to avoid the comparison between components corrupted by noise and achieve accurate recognition, even on highly noisy images. Our experiments were performed in three widely used iris image databases (third version of CASIA, ICE and UBIRIS) and led us to observe that our proposal significantly decreases the error rates in the recognition of noisy iris images.