The first participant presentation sessions (Monday to Wednesday) consisted in volunteers giving 15-minute presentations of their current research work. The remaining sessions (Wednesday to Friday) were for the participants to freely discuss their research with each other. The participants were encouraged to bring as much material as they wanted, regardless if they gave a presentation or not. We provided projectors and panels for posters as needed.
The schedule for the 15-minute presentations is given below, with further details of each one.
Severino Paulo Gomes Neto
Departamento de Informática e Matemática Aplicada, Universidade Federal do Rio Grande do Norte
This project aims at integrating algorithms from Unmanned Navigation and 3D Reconstruction to automate 3D model acquisition. Possible applications of these unmanned systems are in oil prospecting, mining, structural inspection and several others under high-risk conditions. Since the ultimate goal is to use those algorithms in vehicles and robots working in real-time, some algorithms will probably be implemented in GPUs.
The project is currently branched in the following tasks: Describe the Unmanned Navigation taxonomy, Develop an Unmanned Navigation Simulator, Speed up motion hypotheses evaluation, Unfold a scheme for selecting best parameters' values in BRUMA function, Enhance motion parallax estimation and triangulation processes and Develop a scale estimation scheme.
Taha Ridene
Robotics Center, MINES ParisTech
The development of 3D mapping databases is a matter of increasing interest. Databases have recently been developed at different scales and to meet different needs. Such development has been made possible by the implementation of efficient 3D mapping technologies. 3D mapping strategies are based on multi-sensor data fusion usually performed after a preprocessing step that includes registration and filtering. In this presentation, we show our work on registration methods applied to solve problems in 3D urban environment representations issued from a Mobile Mapping System. We improve the stability of convergence, the computation time and handle heterogeneous data sets in various scenarios. The evaluation of 3D processing results is an open problem. In this term, we propose a new evaluation method of 3D registration processing based on binary and fuzzy comparison of geometric features.
Victor Hugo Arroyo Dominguez
Instituto Nacional de Astrofísica, Óptica y Electrónica
3D recovery of metallic specular surfaces using a polarimetric stereo technique
Stereo image analysis is based on establishing correspondences between a pair of images by determining similarity measures for potentially corresponding image parts. Such similarity criteria are only strictly valid for surfaces with Lambertian reflectance characteristics. Specular reflections are viewpoint dependent and may thus cause large intensity differences at corresponding image points. In the presence of specular reflections, traditional stereo approaches are often unable to establish correspondences at all, or the inferred disparity values tend to be inaccurate, or the established correspondences do not belong to the same physical surface point.
Polarization imaging is a powerful tool to observe hidden information from an observed object. It has significant advantages, such as computational efficiency and can be easily applied by adding a polarizer in front of a camera. Although many researchers used polarization in various areas of computer vision, such as object recognition, segmentation and so on, there is very little research in stereo vision based on polarization.
Polarization is a physical intrinsic feature of light which is as important as light intensity. This is an optical phenomena in nature which contains additional information and provides richer description of a scene. The use of polarization, for image understanding, is an enhanced method to sense light parameters from a scene which reveals more significant information than intensity and color.
In this research work, we studied and develop a method that combines stereo and polarization imaging. Considering that one of the main focuses of research in the estereo image analysis area is to get accurate stereo correspondences, we propose a stereo matching algorithm combining the polarization information and the stereo system information.
Lucas Enrique Guaycochea
Facultad de Ingeniería, Universidad de Buenos Aires
This project is a study of the evolution and the current state-of-the-art of the terrain rendering techniques. A solution will be proposed taking advantage of the graphics processing units (GPUs), as their new capabilities refers.
The research is based on the well-known Geometry Clipmaps Technique and will have combined some features taken from newer solutions. As previously introduced, the properties and the performance of the last graphic pipeline model in the GPUs, which is called Shader Model 4.0 and has a new programmable stage, will be analyzed.
Finally, these are the basis from which the solution will arise.
Carlos Alberto Fraga Pimentel Filho
Laboratório de Processamento de Imagens, Universidade Federal de Minas Gerais
The work is concerned with video indexing and retrieval based on visual features. It puts forth an approach for the automatic summary and indexing of digital videos in order to support queries based on visual content within the indexed video's repository. The proposed approach was applied to a database containing more than 34 hours of broadcast news videos. Visual features extracted from the summarized version of the videos were then used for video content indexing. That provided us with the basis for various experiments and analysis on the retrieval of visual content with the application of various techniques implemented in this work. The approach proposes a method for key frame extraction that summarizes video content in a static storyboard, specifically projectec, for key frame retrieval and video access. Thus, the selected key frames are processed in order to extract statistical features as well as wavelet coefficients to represent the video's essence in a very short amount of data while preserving its main content characteristics.
William Robson Schwartz
Department of Computer Science, University of Maryland
Analysis of images involving humans is of significant interest in computer vision because problems such as detection, modeling, recognition, and tracking are fundamental to model interactions between people and understand high-level activities. Visual information contained in images is generally represented using descriptors (features). Many general classes of descriptors have been proposed focusing on different characteristics of images. Therefore, if one considers only a single descriptor, one might ignore useful information for a given task, compromising performance. In this research we consider a rich set of image descriptors analyzed by a statistical technique known as Partial Least Squares (PLS). PLS is a class of methods for modeling relations between sets of observations by means of latent variables and it is used to project exemplars from a very high dimensional feature space onto a low dimensional subspace.
We demonstrate the effectiveness of combining a richer set of descriptors using PLS in two significant tasks in computer vision. First, we propose a method to detect pedestrians. Second, an object recognition framework based on a one-against-all scheme is exploited for face identification.
Click here to download the demonstration video.
Danilo Medeiros Eler
Instituto de Ciências Matemáticas e de Computação, Universidade de São Paulo
Multidimensional Visualization techniques are invaluable tools for analysis of structured and unstructured data with variable dimensionality. This work introduces PEx-Image — Projection Explorer for Images — a tool aimed at supporting analysis of image collections. The tool supports a methodology that employs interactive visualizations to aid user-driven feature detection and classification tasks, thus offering improved analysis and exploration capabilities. The visual mappings employ similarity-based multidimensional projections and point placement to layout the data on a plane for visual exploration. In addition to its application to image databases, we also illustrate how the proposed approach can be successfully employed in simultaneous analysis of different data types, such as text and images, offering a common visual representation for data expressed in different modalities.
Marianela Parodi
Facultad de Ciencias Exactas, Ingeniería y Agrimensura, Universidad Nacional de Rosario
In this presentation, new results on offline signature verification are presented. In particular, a new feature extraction approach based on a circular grid is introduced. Graphometric features used in the rectangular grid segmentation approach are adapted to this new grid geometry. A Support Vector Machine (SVM) based classifier scheme is used for classification tasks and a comparison between the rectangular and the circular grid approaches is performed. The presented results show improvements when using the proposed features with respect to the case of using the features extracted from rectangular grids, specially, in discriminating simple and skilled forgeries.
Denise Guliato
Faculdade de Computação, Universidade Federal de Uberlândia
Fuzzy rule-based systems have been successfully used in the solution of various control problems. A fuzzy classifier takes into account the uncertainty inherent to the major real classification problems. The fuzzy rules, for these systems, can be derived from human experts as linguistic if-then rules. However, in many applications the knowledge required to derive these rules may not be easily available, and humans may be unable to extract the knowledge out of a massive amount of numerical data.
Recently several works have been proposed to automatically generate fuzzy rules from numerical data. Considerable efforts have been concentrated in the use of GA to obtain fuzzy rules automatically and to tune fuzzy partitions of the input space. Genetic algorithms are robust due to the global searching, however involve intensive computation and the results are strongly dependent on the fitness functions and the GA parameters such as number of generations, population size, crossover and mutation rates, tournament size, crossover type and the stop criterium.
The use of rough set to support fuzzy rule-based systems is still a challenge. Few works have been proposed to address the classification problem based on rules using rough sets, however, the rough set theory is not used directly to generate the fuzzy rules. These methods do not take into account the ambiguity of the data, and consequently the lack of evidence (or ignorance) in classifying a given pattern into one of existing classes.
The present work overcomes this problem. We propose a novel method to generate automatically fuzzy rules based on rough set theory. The universe of the knowledge of a given application is grouped into different combinations, with different sizes, using a data mining algorithm. Each one of these groups is referred as granule of the knowledge. The proposed method analyzes each granule of the knowledge, in order to obtain concise rules (reduced number of antecedent terms) with high covered. Due to the lack of information or the uncertainty inherent to the application, two objets can present similar features but belong to different classes. In the face of ambiguous information, the proposed fuzzy classification system is able to distinguish between evidence and ignorance about a given pattern.
The resulting classifier system based on the set of fuzzy rules was tested with the public databases: Iris, Wine, and Wdbc datasets, presenting accuracy rates of 100%, 100%, and 99%, respectively.
Thiago Vallin Spina
Instituto de Computação, Universidade Estadual de Campinas
Among the many challenges in interactive editing of natural images and videos lies the use of effective techniques for composing segmented regions of interest (e.g., people, objects, animals). This research deals with this problem using image processing tools based on the image foresting transform (IFT). IFT is a methodology for the development of image processing operators based on connectivity. It is able to reduce the segmentation process to the choice of a few seed pixels on the image (user-drawn strokes, for example). The composition of a segmented foreground and a new background (alpha matting) is a more general problem then the binary segmentation, and shall be addressed using IFT-based pattern classifiers. We are also investigating means to combine region-based and boundary-based segmentation paradigms (namely the differential IFT and live-wire-on-the-fly) for more effective result. Finally, the developed framework will be extended to video.
Cesar Castelo Fernandez
Instituto de Computação, Universidade Estadual de Campinas
The main goal of this project is to improve existing techniques for pattern recognition, allowing faster and more accurate classification. We are using a classifier based on graphs, the Optimum-Path Forest classifier, which is so fast and accurate, being that our first publication is about a method to improve its accuracy, presented as a general-purpose classifier.
In the next stage of the project we will research about methods for reducing the training and testing time. These new methods will be based in mixing the OPF with other techniques which have their own vantages, like SVM, Neural Networks and other variants of the OPF itself. The goal of this stage is to obtain a hybrid classifier with the best features from the others. For example, it could perform training with the OPF algorithm, and testing with SVM. This new method will be presented as a general- purpose classifier too, having the main focus of being applied in real-time classification of large datasets.
Finally, the methods developed will be applied for Interactive Segmentation of Medical Images, which is a kind of application that needs fast classification methods since the images used are so big, representing great challenges.
Eduardo Alvarez Ribeiro
Faculdade de Medicina de Ribeirão Preto, Universidade de São Paulo
A few recent studies have proposed computed-aided methods for the detection and analysis of vertebral bodies in radiographic images. This paper presents a method based on Gabor filters. Forty-one lateral lumbar spinal X-ray images from different patients were included in the study. For each image, a radiologist manually delineated the vertebral plateaus of L1, L2, L3, and L4 using a software tool for image display and mark-up. Each original image was filtered with a bank of 180 Gabor filters. The angle of the Gabor filter with the highest response at each pixel was used to derive a measure of the strength of orientation or alignment. In order to limit the spatial extent of the image data and the derived features in further analysis, a semi-automated procedure was applied to the original image. A neural network utilizing the logistic sigmoid function was trained with pixel intensity from the original image, the result of manual delineation of the plateaus, the Gabor magnitude response, and the alignment image. The average overlap between the results of detection by image processing and manual delineation of the plateaus of L1-L4 in the 41 images tested was 0.917. The results are expected to be useful in the analysis of vertebral deformities and fractures.
Mariela Azul Gonzalez
Departamento de Electrónica, Universidad Nacional de Mar del Plata
PhD. Thesis: Marker definition for Watershed Transform using Mathematical Morphology Operators. We developed algorithms combining mathematical morphology, clustering and fuzzy logic for segmentation of digital images of bone marrow biopsies.
Current Work: I am currently working on characterization and quantification of shape, size and direction of objects in 2D Medical Images. The algorithms are being applied to the processing of images of growing neurons and polymeric scaffolds, generated by tissue engineering techniques. Future work is going to be oriented to 3D Medical Image processing.
Matheus Cardoso Moraes
Escola Politécnica, Universidade de São Paulo
Intravascular Ultrasound (IVUS) segmentation gives more precise vessel information, leading to better analyses. The goal herein is to segment the media-adventitia border in IVUS images by combining a set of imaging processing operations: Speckle Reducing Anisotropic Diffusion (SRAD); Intensity Transformation; Wavelet; Otsu and Mathematical Morphology. The segmentation accuracy was assessed by segmenting 100 challenging images and comparing the result to their corresponding Gold Standard, it was obtained an average of True Positive (TP(%)) = 92.83±4.91, False Positive (FP(%)) = 3.43±3.47, False Negative (FN(%)) = 7.17±4.91, Max False Positive (MaxFP (mm)) = 0.27±0.22, Max False Negative (MaxFN(mm)) = 0.31±0.2. Likewise, by comparing the outcome with a recent one in the literature, its effectiveness is corroborated.
Wen Li
Department of Biomedical Engineering, University of Iowa
This research is about human cerebral cortex analysis on images of brain MRI. The goal is to take routine MRI images (T1/T2/Protein weighted) and do the parcellation on cortical surfaces automatically. We have developed a fully automated pipeline for surface generation and parcellation using open source tools. Currently, we are trying to incorporate the whole pipeline into the newest version of BRAINS (Brain Research: Analysis of Images, Networks and Systems) image analysis software. Spherical Demons Registration was used to register the atlas surface, as a moving mesh to the subject surface, as a fixed mesh in spherical domain. And labels on the atlas surface were warped onto the subject surface after the registration. The method has been tested on six subjects, one of which was used as an atlas. The overlapping between the warped labels and manual labels was calculated by Dice Index. The average value of Dice Index among five results was 0.88.
Carlos Fernando Crispim Junior
Instituto de Engenharia Biomédica, Universidade Federal de Santa Catarina
In my doctorate project, I examine image and descriptive statistic attributes that correctly represent the characteristics of video images of behavioral events in laboratory animals (e.g.: rats, mice). The images came from digital movie recordings taken from laboratory animals during pharmacological experiments. Behavioral descriptors are pre-evaluated through parametric or non-parametric statistical tests, and then are tested as input patterns for artificial neural networks and support vector machines. The study of techniques that describe behavioral events will improve the design of computer-aided diagnosis tools (based in video recordings) used in pharmacology and ethology. Furthermore, the study of descriptors for complex signals with non-normal distribution and strong time dependency will be helpful for other computer science applications.
Filip Malmberg
Centre for Image Analysis, Uppsala University
To make precise measurements of object features (e.g., area and perimeter) based on digital images is at the core of almost any image analysis application. Such measurements often require accurate segmentation of relevant objects in the image. However, even under the assumption that a correct segmentation is known, the measurement accuracy is still limited by the fact that we are trying to estimate features of continuous (real-world) objects based on a discrete, sampled, representation of the objects.
In this talk, I will present some recent work on formulating a graph-based framework for sub-pixel segmentation. Experiments, performed on both synthetic and real objects, indicate that the proposed framework facilitates significant improvements in precision and accuracy of feature estimates.
Francisco Paulo Magalhães Simões
Centro de Informática, Universidade Federal de Pernambuco
Augmented Reality (AR) systems support the coexistence of virtual and real worlds. This coexistence is made possible through environment tracking in order to register synthetic elements accordingly. Optical tracking is often used for this purpose due to cost, accuracy and robustness requirements. In model based optical tracking there are two major categories: Texture based an edge based techniques. This project is focused on development of model based markerless augmented reality techniques, more precisely in a texture based technnique with the use of interest points. As interest points can suffer from ambiguities that can lead to wrong matches and tracking failure, the proposed technique uses a particle filter approach to calculate the new camera pose and achieve a better tracking result.
Teodoro Schmidt
GRIMA - Grupo de Inteligencia de Máquinas, Pontificia Universidad Católica de Chile
Biometric technologies have always been the primary tools for certifying the identity of any individual. But the cost of sensing hardware plus the degree of physical invasion that it's required to obtain a reasonable classification success have been considered the two major drawbacks of these technologies.
Nevertheless, there is one means of identification of individuals that it's generally accepted by most societies and in which even the most critical transactions of today's world rely on, and that is the signature of a person.
We present a novel approach on Signature Recognition proposing the use of face recognition algorithms to obtain class descriptors and the use of a simple classifier to recognize a person's signature.
Our proposal states algorithms that dynamically detect the geometry and sense of signature, applying a linear transformation that takes the data representation into gray scale, then processing the images applying Principal Components Analysis (PCA) and Linear Discriminant Analysis (LDA), thus determining homologous descriptors to EigenFaces and FisherFaces, which we called EigenSignatures and FisherSignatures. These descriptors are then classified using a simple KNN classifier.
So far our experimental results have revealed that the best rate of performance is obtained using LDA, with 94% of accuracy, which puts this technique close the the most successful biometric technologies. Finally, the overall implication and novelty of this technique is that it allows the storage of additional information aside from geometry, strengthening the recognition of signatures through the extraction of a new type of characteristic (feature), applying transformations that allow the processing of data as gray scale images, all performed without the need of extra additional hardware, aside from a cheap tablet-like device.