Pablo Suau

Information Theory in Computer Vision and Pattern Recognition

Information theory has proved to be effective for solving many computer visionand pattern recognition (CVPR) problems (such as image matching, clustering and segmentation, saliency detection, feature selection, optimal classifier design and many others). Nowadays, researchers are widely bringing information theory elements to the CVPR arena. Among these elements there are measures (entropy, mutual information), principles (maximum entropy, minimax entropy) and theories (rate distortion theory, method of types). This book explores and introduces the latter elements through an incremental complexity approach at the same time where CVPR problems are formulated and the most representative algorithms are presented. Interesting connections between information theory principles when applied to different problems are highlighted, seeking a comprehensive research roadmap. The result is a novel tool both for CVPR and machine learning researchers, and contributes to a cross-fertilization of both areas.

Constellations and the unsupervised learning of graphs

In this paper, we propose a novel method for the unsupervised clustering of graphs in the context of the constellation approach to object recognition. Such method is an EM central clustering algorithm which builds prototypical graphs on the basis of fast matching with graph transformations. Our experiments, both with random graphs and in realistic situations (visual localization), show that our prototypes improve the set median graphs and also the prototypes derived from our previous incremental method. We also discuss how the method scales with a growing number of images.

Bayesian optimization of the scale saliency filter

The scale saliency feature extraction algorithm by Kadir and Brady has been widely used in many computer vision applications. However, when compared to other feature extractors, its computational cost is high. In this paper, we analyze how saliency evolves through scale space, demonstrating an intuitive idea: if an image region is homogeneous at higher scales, it will probably also be homogeneous at lower scales. From the results of this analysis we propose a Bayesian filter based on Information Theory, that given some statistical knowledge about the images being considered, discards pixels from an image before applying the scale saliency detector. Experiments show that if our filter is used, the efficiency of the original algorithm increases with low localization and detection error.

Contextual visual localization: cascaded submap classification, optimized saliency detection, and fast view matching

In this paper, we present a novel coarse-to-fine visual localization approach: contextual visual localization. This approach relies on three elements: (i) a minimal-complexity classifier for performing fast coarse localization (submap classification); (ii) an optimized saliency detector which exploits the visual statistics of the submap; and (iii) a fast view-matching algorithm which filters initial matchings with a structural criterion. The latter algorithm yields fine localization. Our experiments show that these elements have been successfully integrated for solving the global localization problem. Context, that is, the awareness of being in a particular submap, is defined by a supervised classifier tuned for a minimal set of features. Visual context is exploited both for tuning (optimizing) the saliency detection process, and to select potential matching views in the visual database, close enough to the query view.

Analysis of the schrödinger operator in the context of graph characterization

In this paper, we apply the solution of the Schrodinger equation, i.e. the Schrodinger operator, to the graph characterization problem. The motivation behind this approach is two-fold. Firstly, the mathematically similar heat kernel has been used in the past for this same problem. And secondly, due to the quantum nature of the Schrodinger equation, our hypothesis is that it may be capable of providing richer sources of information. The two main features of the Schrodinger operator that we exploit in this paper are its non-ergodicity and the presence of quantum interferences due to the existence of complex amplitudes with both positive and negative components. Our proposed graph characterization approach is based on the Fourier analysis of the quantum equivalent of the heat flow trace, thus relating frequency to structure. Our experiments, performed both on synthetic and real-world data, demonstrate that this new method can be successfully applied to the characterization of different types of graph structures.

Region and constellations based categorization of images with unsupervised graph learning

In this paper, we address the problem of image categorization with a fast novel method based on the unsupervised clustering of graphs in the context of both region-based segmentation and the constellation approach to object recognition. Such method is an EM central clustering algorithm which builds prototypical graphs on the basis of either Softassign or fast matching with graph transformations. We present two realistic applications and their experimental results: categorization of image segmentations and visual localization. We compare our graph prototypes with the set median graphs. Our results reveal that, on the one hand, structure extracted from images improves appearance-based visual localization accuracy. On the other hand, we show that the cost of our central graph clustering algorithm is the cost of a pairwise algorithm. We also discuss how the method scales with an increasing amount of images. In addition, we address the scientific question of what are the bounds of structural learning for categorization. Our in-depth experiments both for region-based and feature-based image categorization, will show that such bounds depend hardly on structural variability.

Robust artificial landmark recognition using polar histograms

New results on our artificial landmark recognition approach are presented, as well as new experiments in order to demonstrate the robustness of our method. The objective of our work is the localization and recognition of artificial landmarks to help in the navigation of a mobile robot. Recognition is based on interpretation of histograms obtained from polar coordinates of the landmark symbol. Experiments prove that our approach is fast and robust even if the database has an high number of landmarks to compare with.

Graph Characteristics from the Schrödinger Operator

In this paper, we show how the Schrodinger operator may be applied to the problem of graph characterization. The motivation is the similarity of the Schrodinger equation to the heat difussion equation, and the fact that the heat kernel has been used in the past for graph characterization. Our hypothesis is that due to the quantum nature of the Schrodinger operator, it may be capable of providing richer sources of information than the heat kernel. Specifically the possibility of complex amplitudes with both negative and positive components, allows quantum interferences which strongly reflect symmetry patterns in graph structure. We propose a graph characterization based on the Fourier analysis of the quantum equivalent of the heat flow trace. Our experiments demonstrate that this new method can be succesfully applied to characterize different types of graph structures.

Adapting hausdorff metrics to face detection systems: a scale-normalized Hausdorff distance approach

Template matching face detection systems are used very often as a previous step in several biometric applications. These biometric applications, like face recognition or video surveillance systems, need the face detection step to be efficient and robust enough to achieve better results. One of many template matching face detection methods uses Hausdorff distance in order to search the part of the image more similar to a face. Although Hausdorff distance involves very accurate results and low error rates, overall robustness can be increased if we adapt it to our concrete application. In this paper we show how to adjust Hausdorff metrics to face detection systems, presenting a scale-normalized Hausdorff distance based face detection system. Experiments show that our approach can perform an accurate face detection even with complex background or varying light conditions.

Bypass information-theoretic shape similarity from non-rigid points-based alignment

In this paper we present several information-theoretic similiarity measures for shape retrieval in combination with non-rigid registration processes. The challenging property of these measures is that they are bypass divergences, that is, do not require the estimation of the probability density function for each shape. After presenting the dissimilarities and proposing some new ones, we analyze their performance in terms of average recall for a very difficult database (GatorBait) with many classes, few examples and high degree of intra-class variability. We also test these measures in a subset of the the well known MPEG7 part B database. Our experiments show that the Henze-Penrose divergence outperforms the other ones in 2D shape retrieval. We uncover also very competitive and more efficient measures in both cases.