Catálogo de publicaciones - libros

The modelisation of human faces from images can be done by the mean of morphable models such as AAMs. However, fitting such models without previous estimations is a challenging task. Shape estimation needs a close texture reference, and texture approximation requires shape knowledge. In this paper, we address the efficiency of sampling and generic encoding in regard to the shape alignment accuracy, without previous texture approximation. The hybrid method we propose is based on a relative barycentric resampling of the face model, a generic coding of the reference texture and a normalized cost function. We also present a new warping function definition to simplify the initial global parameter estimation. These new subsampling and encoding frameworks improve the accuracy of facial shape alignment in unconstrained cases.

Face recognition represents a challenging research topic which has been investigated by means of many techniques operating either in 2D or 3D, and, more recently, even through multi-modal approaches. Whatever the methodology used to compare any two faces, the main concern has been on recognition accuracy, often disregarding the efficiency issue which may be crucial in a large scale one-to-many recognition application. This paper presents a Graphic Processing Unit (GPU) assisted face recognition method, operating on 4D data (geometry + texture). It exploits augmented normal map, a 32 bit deep color bitmap, as face descriptor, allowing ultra fast face comparison through the specialized hardware (pixel shaders) available in almost any recently designed PC graphic boards. The proposed approach addresses facial expression changes and presence of beard by means of two (subject specific) filtering masks. We include preliminary experimental results on a large gallery of faces.

The aim of the present work is the recognition of human face visual information, in order to automatically control the access to restricted areas, granting access to authorized “clients” and barring the entrance to “impostors”. The vision system assembled performed the image acquisition, processing and recognition by first creating a database with a single view of each “client” and then by using multiple test images of each individual candidate to access. To get the test images, a video sequence was captured during the individual’s approach path to the camera. Because subjects presented themselves in a random pose before the camera, the synthesis of frontal views was incorporated, by using a view-morphing method. The modelling and the recognition were handled through the use of ICA methods. The identification of valid “clients” was fully successful. In order to check the rejection of “impostors”, a leave-one-out test was performed which gave promising results.

This paper exploits the properties of the commute time for the purposes of graph matching. Our starting point is the lazy random walk on the graph, which is determined by the heat-kernel of the graph and can be computed from the spectrum of the graph Laplacian. We characterise the random walk using the commute time between nodes, and show how this quantity may be computed from the Laplacian spectrum using the discrete Green’s function. We use the commute-time to locate the minimum spanning tree of the graph. The spanning trees located using commute time prove to be stable to structural variations. We match the graphs by applying a tree-matching method to the spanning trees. We experiment with the method on synthetic and real-world image data, where it proves to to be effective.

In this paper we propose an efficient algorithm to perform a polynomial approximation of the vector field derived from the usual distance mapping method. The main ingredients consist of minimizing a quadratic functional and transforming this problem in an appropriate setting for implementation. With this approach, we reduce the problem of obtaining an approximating polynomial vector field to the resolution of a not expansive linear algebraic system. By this procedure, we obtain an analytical shape representation that relies only on some coefficients. Fidelity and numerical efficiency of our approach are presented on illustrative examples.

This paper presents a template matching technique to identify the location and orientation of an object by a fast algorithm. The fundamental principle in template matching is to minimize a potential energy function, which is a quantitative representation of the ’closeness’ of a defined object (template) relative to a portion of an image. However, the computation of potential energy suffers a major drawback in efficiency. A significant amount of the processing time is dedicated to match the information from the template to the image. This work proposes an alternative way to match the template and the image that reduces the number of operations from O ( nm ) to O ( n ) in calculating the potential energy of a template and an image that have n and m number of edge pixels, respectively. This work illustrates this approach by template edge matching which uses the edge information to perform the template matching. The experimental results show that while the proposed method produces a slightly larger error in the resulting template location, the processing time is decreased by a factor of 4.8 on average.

Shape recognition methods are often based on feature comparison. When features are of different natures, combining the value of distances or (dis-)similarity measures is not easy since each feature has its own amount of variability. Statistical models are therefore needed. This article proposes a statistical method, namely an a contrario method, to merge features derived from several families of size functions. This merging is usually achieved through a touchy normalizing of the distances. The proposed model consists in building a probability measure. It leads to a global shape recognition method dedicated to perceptual similarities.

This paper deals with designing an algorithm for feature pairing in vision, based on the ”stable marriages” paradigm. Our SZ is an extension of the recently published BZ algorithm. BZ scans the so-called ”marriage table” to optimize global satisfaction and equity over all couples. It still gets about 5% unstable results in average. After a case study that sorts blocking situations into 4 types, we explain here how to resolve unstability in forcing blocking pairs to marry wrt. their type. SZ is compared to BZ and Gale-Shapley on 40000 instances of a 200 persons large population. An example of stereo reconstruction by SZ is given for illustration.

Affine point pattern matching (APPM) is an integral part of many pattern recognition problems. Given two sets P and Q of points with unknown assignments p _ i → q _ j between the points, no additional information is available. The following task must be solved: – Find an affine transformation T such that the distance between P and the transformed set Q ′= T Q is minimal. In this paper, we present a new approach to the APPM problem based on matching in bipartite graphs. We have proved that the minimum of a cost function is an invariant under special affine transformations. We have developed a new algorithm based on this property. Finally, we have tested the performance of the algorithm on both synthetically generated point sets and point sets extracted from real images.

In this work, a geometric invariant curve and surface normalization method is presented. Translation, scale and shear are normalized by Principal Component Analysis (PCA) whitening. Independent Component Analysis (ICA) and the third order moments are then employed for rotation and reflection normalization. By applying this normalization, curves and surfaces that are related by geometric transformations (affine or rigid) can be transformed into a canonical representation. Proposed technique is verified with several 2D and 3D object matching and recognition experiments.