Catálogo de publicaciones - libros

We propose to combine simple discriminators for object discrimination under the maximum entropy framework or equivalently under the maximum likelihood framework for the exponential family. The duality between the maximum entropy framework and maximum likelihood framework allows us to relate two selection criteria for the discriminators that were proposed in the literature. We illustrate our approach by combining nearest prototype discriminators that are simple to implement and widely applicable as they can be constructed in any feature space with a distance function. For efficient run-time performance we adapt the work on “alternating trees” for multi-class discrimination tasks. We report results on a multi-class discrimination task in which significant gains in performance are seen by combining discriminators under our framework from a variety of easy to construct feature spaces.

The problem of searching for a model-based scene interpretation is analyzed within a probabilistic framework. Object models are formulated as generative models for range data of the scene. A new statistical criterion, the truncated object probability, is introduced to infer an optimal sequence of object hypotheses to be evaluated for their match to the data. The truncated probability is partly determined by prior knowledge of the objects and partly learned from data. Some experiments on sequence quality and object segmentation and recognition from stereo data are presented. The article recovers classic concepts from object recognition (grouping, geometric hashing, alignment) from the probabilistic perspective and adds insight into the optimal ordering of object hypotheses for evaluation. Moreover, it introduces point-relation densities, a key component of the truncated probability, as statistical models of local surface shape.

The aim of this work is to propose an adaptation of optimal path based interactive tools for image segmentation (related to [] and [] approaches). We efficiently use both discrete [] and continuous [] path search approaches. The segmentation relies on the notion of energy function and we introduce the possibility of complete adaptation of each individual energy term, as well as of their relative weights. Non-specialist users have then a full control of the drawing process which automatically selects the most relevant set of features to steer the path extraction. Tests have been performed on a large variety of medical images.

This paper investigates a technique to extend the tree edit distance framework to allow the simultaneous matching of multiple tree structures. This approach extends a previous result that showed the edit distance between two trees is completely determined by the maximum tree obtained from both tree with node removal operations only. In our approach we seek the minimum structure from which we can obtain the original trees with removal operations. This structure has the added advantage that it can be extended to more than two trees and it imposes consistency on node matches throughout the matched trees. Furthermore through this structure we can get a “natural” embedding space of tree structures that can be used to analyze how tree representations vary in our problem domain.

Over the last decade, a number of methods for geometric matching based on a branch-and-bound approach have been proposed. Such algorithms work by recursively subdividing transformation space and bounding the quality of match over each subdivision. No direct comparison of the major implementation strategies has been made so far, so it has been unclear what the relative performance of the different approaches is. This paper examines experimentally the relative performance of different implementation choices in the implementation of branch-and-bound algorithms for geometric matching: alternatives for the computation of upper bounds across a collection of features, and alternatives the order in which search nodes are expanded. Two major approaches to computing the bounds have been proposed: the matchlist based approach, and approaches based on point location data structures. A second issue that is addressed in the paper is the question of search strategy; branch-and-bound algorithms traditionally use a “best-first” search strategy, but a “depth-first” strategy is a plausible alternative. These alternative implementations are compared on an easily reproducible and commonly used class of test problems, a statistical model of feature distributions and matching within the COIL-20 image database. The experimental results show that matchlist based approaches outperform point location based approaches on common tasks. The paper also shows that a depth-first approach to matching results in a 50-200 fold reduction in memory usage with only a small increase in running time. Since matchlist-based approaches are significantly easier to implement and can easily cope with a much wider variety of feature types and error bounds that point location based approaches, they should probably the primary implementation strategy for branch-and-bound based methods for geometric matching.

We address the problem of face recognition from a large set of images obtained over time - a task arising in many surveillance and authentication applications. A set or a sequence of images provides information about the variability in the appearance of the face which can be used for more robust recognition. We discuss different approaches to the use of this information, and show that when cast as a statistical hypothesis testing problem, the classification task leads naturally to an information-theoretic algorithm that classifies sets of images using the relative entropy (Kullback-Leibler divergence) between the estimated density of the input set and that of stored collections of images for each class. We demonstrate the performance of the proposed algorithm on two medium-sized data sets of approximately frontal face images, and describe an application of the method as part of a view-independent recognition system.

We present a method — termed Helmholtz stereopsis — for reconstructing the geometry of objects from a collection of images. Unlike most existing methods for surface reconstruction (e.g., stereo vision, structure from motion, photometric stereo), Helmholtz stereopsis makes no assumptions about the nature of the bidirectional reflectance distribution functions (BRDFs) of objects. This new method of multinocular stereopsis exploits Helmholtz reciprocity by choosing pairs of light source and camera positions that guarantee that the ratio of the emitted radiance to the incident irradiance is the same for corresponding points in the two images. The method provides direct estimates of both depth and field of surface normals, and consequently weds the advantages of both conventional and photometric stereopsis. Results from our implementations lend empirical support to our technique.

Determining shape from stereo has often been posed as a global minimization problem. Once formulated, the minimization problems are then solved with a variety of algorithmic approaches. These approaches include techniques such as dynamic programming min-cut and alpha-expansion. In this paper we show how an algorithmic technique that constructs a discrete spatial minimal cost surface can be brought to bear on stereo global minimization problems. This problem can then be reduced to a single min-cut problem. We use this approach to solve a new global minimization problem that naturally arises when solving for three-camera (trinocular) stereo. Our formulation treats the three cameras symmetrically, while imposing a natural occlusion cost and uniqueness constraint.

Stereo matching is an ill-posed problem for at least two principal reasons: (1) because of the random nature of match similarity measure and (2) because of structural ambiguity due to repetitive patterns. Both ambiguities require the problem to be posed in the regularization framework. Continuity is a natural choice for a prior model. But this model may fail in low signal-to-noise ratio regions. The resulting artefacts may then completely spoil the subsequent visual task.

A question arises whether one could (1) find the unambiguous component of matching and, simultaneously, (2) identify the ambiguous component of the solution and then, optionally, (3) regularize the task for the ambiguous component only. Some authors have already taken this view. In this paper we define a new which is a condition a set of matches must satisfy to be considered unambiguous at a given confidence level. It turns out that for a given matching problem this set is (1) unique and (2) it is already a matching. We give a fast algorithm that is able to find the largest stable matching. The algorithm is then used to show on real scenes that the unambiguous component is quite dense (10–80%) and error-free (total error rate of 0.3–1.4%), both depending on the confidence level chosen.