Catálogo de publicaciones - libros

In this paper, we address the problem of automating the partial representation from real world data with an unknown a priori structure. Such representation could be very useful for the further construction of an automatic hierarchical data model. We propose a three stage process using data normalisation and the data intrinsic dimensionality estimation as the first step. The second stage uses a modified sparse Non-negative matrix factorization (sparse NMF) algorithm to perform the initial segmentation. At the final stage region growing algorithm is applied to construct a mask of the original data. Our algorithm has a very broad range of a potential applications, we illustrate this versatility by applying the algorithm to several dissimilar data sets.

Single photon emission computed tomography (SPECT) of dopamine transporters with Tc-TRODAT-1 has recently been proposed to provide valuable information of assessing the dopaminergic system. In order to measure the binding ratio of the nuclear medicine, registering magnetic resonance imaging (MRI) and SPECT image is a significant process. Therefore, an automated MRI/SPECT image registration algorithm of using an adaptive similarity metric is proposed. This similarity metric combines anatomic features characterized by specific binding (SB), the mean counts per voxel within the specific tissues, of nuclear medicine and distribution of image intensity characterized by the Normalized Mutual Information (NMI). In addition, we have also built a computer-aid clinical diagnosis system which automates all the processes of MRI/SPECT registration for further evaluation of Parkinson’s disease. Clinical MRI/SPECT data from eighteen healthy subjects and thirteen patients are involved to validate the performance of the proposed system. Comparing with the conventional NMI-based registration algorithm, our system reduces the target of registration error (TRE) from >7 mm to approximate 4 mm. From the view point of clinical evaluation, the error of binding ratio, the ratio of specific-to-non-specific Tc-TRODAT-1 binding, is 0.20 in the healthy group and 0.13 in the patient group via the proposed system.

We present a gesture recognition system for recognizing hand movements in near realtime. The system uses a infra-red time-of-flight range camera with up to 30 Hz framerate to measure 3d-surface points captured from the hand of the user. The measured data is transformed into a cloud of 3d-points after depth keying and suppression of camera noise by median filtering. Principle component analysis (PCA) is used to obtain a first crude estimate on the location and orientation of the hand. An articulated hand model is fitted to the data to refine the first estimate. The unoptimized system is able to estimate the first 7 Degrees of Freedom of the hand within 200 ms. The reconstructed hand is visualized in AVANGO/Performer and can be utilized to implement a natural man-machine interface. The work reviews relevant publications, underlines the advantages and shortcomings of the approach and provides an outlook on future improvements.

In this paper we take a fresh look at the problem of extracting shape from contours of human faces. We focus on two key questions: how can we robustly fit a 3D face model to a given input contour; and, how much information about shape does a single contour image convey.

Our system matches silhouettes and inner contours of a PCA based Morphable Model to an input contour image. We discuss different types of contours in terms of their effect on the continuity and differentiability of related error functions and justify our choices of error function (modified Euclidean Distance Transform) and optimization algorithm (Downhill Simplex).

In a synthetic test setting we explore the limits of accuracy when recovering shape and pose from a single correct input contour and find that pose is much better captured by contours than is shape. In a semi-synthetic test setting – the input images are edges extracted from photorealistic renderings of the PCA model – we investigate the robustness of our method and argue that not all discrepancies between edges and contours can be solved by the fitting process alone.

Nowadays, there are many applications that take advantage of the availability of three-dimensional (3D) data sets. These objects are represented as complex polygonal surfaces formed by hundreds of thousands of polygons, which causes a significant increase in the cost of storage, transmission and visualisation. Multiresolution modeling, which allows an object to be represented by set of approximations, each with a different number of polygons, has been successfully presented as a solution for the efficient manipulation of this type of objects. The main contribution of this work is the use of the complete lifting scheme for the multiresolution analysis of irregular meshes with proposition of new prediction block.

This paper is concerned itself with the analysis of the two-phase Mumford-Shah model also known as the model introduced by Chan and Vese. It consists of approximating an observed image by a piecewise constant image which can take only two values. First we show that this model with the -norm as data fidelity yields a contrast invariant filter which is a well known property of morphological filters. Then we consider a discrete version of the original problem. We show that an inclusion property holds for the minimizers. The latter is used to design an efficient graph-cut based algorithm which computes an exact minimizer. Some preliminary results are presented.

In this paper, we propose an algorithm for plane segmentation using an optical flow field computed from successive images captured by an uncalibrated moving camera. The proposing method does not require any restrictions on the camera motion and the camera-configuration geometry. Our segmentation algorithm is based on the algorithm of dominant-plane detection. The dominant plane is a planar area in the world, and it corresponds to the largest part of an image. By iterative processing dominant-plane detection, our algorithm detects multiple planes in an image. We present experimental results using image sequences observed with a moving camera in a synthesized environment and a real environment.

In this paper, we propose a new gaze estimation method by analyzing the cornea surface model which is estimated through three dimensional analysis of human eye in HMD (Head Mounted Display) environments. This paper has four advantages over previous works. First, in order to obtain accurate gaze position, we use a cornea sphere model based on Gullstrand eye scheme. Second, we calculate the 3D position of the cornea sphere and a gaze vector by using a camera, three collimated IR-LEDs and one illuminated IR-LED. Third, three coordinates such as camera, monitor and eye coordinates are unified, which can simplify the complex 3D converting calculation and allow for calculation of the 3D eye position and gaze position on a HMD monitor. Fourth, a simple user dependent calibration method is proposed by gazing at one position of HMD monitor based on Kappa compensation. Experimental results showed that the average gaze estimation error of the proposed method was 0.89 degrees.

This paper describes expression space generation technology that enables animators to control the expressions of 3-dimensional avatars in real-time by selecting a series of expressions from facial expression space. In this system, approximately 2400 facial expression frames are used to generate facial expression space. In this paper, distance matrixes that present distances between facial characteristic points are used to show the state of an expression. The set of these distance matrixes is defined as facial expression space. However, this facial expression space is not space that can be transferred to one space or another in a straight line, when one expression changes to another. In this technology, the route for moving from one expression to another is approximately inferred from captured facial expression data. First, it is assumed that two expressions are close to each other when the distance between distance matrixes that show facial expression states is below a certain value. When two random facial expression states are connected with the set of a series of adjacent expressions, it is assumed that there is a route between the two expressions. It is further assumed that the shortest path between two facial expressions is the path when one expression moves to the other expression. Dynamic programming is used to find the shortest path between two facial expressions. The facial expression space, which is the set of these distance matrixes, is multidimensional space. The facial expression control of 3-dimensional avatars is carried out in real-time when animators navigate through facial expression space. In order to assist this task, multidimensional scaling is used for visualization in 2-dimensional space, and animators are told to control facial expressions when using this system. This paper evaluates the results of the experiment.

We consider 3D shape description as a probability modeling problem. The local surface properties are first measured via various features, and then the probability density function (pdf) of the multidimensional feature vector becomes the shape descriptor. Our prior work has shown that, for 3D object retrieval, pdf-based schemes can provide descriptors that are computationally efficient and performance-wise on a par with or better than the state-of-the-art methods. In this paper, we specifically focus on discretization problems in the multidimensional feature space, selection of density evaluation points and dimensionality reduction techniques to further improve the performance of our density-based descriptors.