Séminaires

Les séminaires du projet Ariana ont lieu à l'INRIA Sophia Antipolis (plan), la salle ainsi que les résumés (en français ou/et en anglais) étant affichés dès que possible. Si vous le souhaitez, vous pouvez consulter l'agenda des séminaires des années précédentes : 2024, 2023, 2022, 2021, 2020, 2019, 2018, 2017, 2016, 2015, 2014, 2013, 2012, 2011, 2010, 2009, 2008, 2007, 2006, 2005, 2004, 2003, 2002, 2001, 2000, 1999, et 1998.

Anciens séminaires du projet Ariana :

Titre	Intervenant	Date/Lieu	Résumé
Detecting noticeable features in ambiguously registered image views	Tamás SZIRANYI Professeur MTA SZTAKI, Budapest	13/12/2010 14h30 salle Coriolis
Résumé (anglais) : In this talk, I will review our latest results in change detection on aerial image pairs taken with the time differences of several years and in different seasonal conditions; a conditional mixed Markov model and new saliency methods, including Harris function and local active contours have been developed to find changed noticeable features. I will also introduce the problems and some solutions for freely positioned camera/sensor networks: how can swarm optimization solve the connectedness problem of flexible camera networks? What could be the fitness function for registering views in unstructured environment? I will also show that shadow might be beneficial after it has been successfully detected.
Surveillance maritime par analyse d'images satellitaires optiques panchromatiques	Nadia PROIA Chercheur Associé Université des Antilles et de la Guyane	06/12/2010 14h30 salle Coriolis
Résumé (français) : La surveillance maritime constitue aujourd'hui un énorme enjeu dans le domaine de la sécurité et dans celui de l'économie. Nous présenterons une méthode qui s’appuie sur les techniques Bayésiennes classiques de détection. La détection est décomposée en trois étapes. La première consiste en une pré-détection des cibles qui nous fournit des cibles candidates. La deuxième étape est une segmentation précise de chaque cible candidate. Et la troisième phase est une classification des candidates en trois classes : petit bateau, grand bateau ou fausse alarme. Pour compléter cette étude et se replacer dans un contexte industriel et opérationnel, nous présentons des expérimentations de l'ensemble de l'algorithme sur des images complètes pour évaluer les performances et le temps de traitement.
Imagerie biologique quantitative : des cellules, des images et des nombres	Jean-Christophe OLIVO MARIN Directeur de Recherche Institut Pasteur, Paris	29/11/2010 14h30 salle Coriolis
Résumé (français) : L'analyse informatisée des images constitue le seul moyen d'étudier de manière systématique et exhaustive les données brutes issues de la microscopie multidimensionnelle et d'établir un pont entre l'imagerie et l'interprétation expérimentale. Nous présenterons des méthodes pour le traitement et la quantification de séquences temporelles 3D et leur utilisation dans des projets biologiques, en particulier des algorithmes de suivi de particules multiples, de séparation de sources et des modèles de contours actifs. Nous illustrerons les applications de ces méthodes dans des projets liés à la détection et le suivi de virus dans des cellules, la dynamique de particules fluorescentes et la mobilité de parasites.
On-line, semi-supervised, and multiple-instance learning for computer vision	Horst BISCHOF Professeur Université technique de Graz, Autriche	20/09/2010 14h30 salle Galois-Coriolis
Résumé (anglais) : In this talk I will review the progress we made in on-line, semi-supervised and multiple instance learning. I will argue that these three properties are desirable features for a learning algorithm suitable for computer vision tasks. I will present boosting and random forest variants. After introducing the algorithms I will demonstrate the applications to various vision applications, including tracking and object detection.
Classification of Hyperspectral Data Using Spectral-Spatial Approaches	Yuliya TARABALKA Doctorante Université d'Islande, Institut de Technologie de Grenoble	06/08/2010 14h30 salle Euler Violet
Résumé (anglais) : In this talk, novel strategies for spectral-spatial classification of hyperspectral images will be presented. One of the recently proposed approaches consists of performing image segmentation in order to use every region from the segmentation map as an adaptive spatial neighborhood for all the pixels within this region. Different segmentation techniques are investigated and extended to the case of hyperspectral images. We also propose approaches to reduce oversegmentation in an image, which is achieved by automatically marking the spatial structures before performing a marker-controlled segmentation. Our proposal is to analyze probabilistic classification results for selecting the most reliably classified pixels as markers of spatial regions. Then, different approaches for marker-controlled region growing are developed. The new techniques significantly decrease oversegmentation, improve classification accuracies and provide classification maps with more homogeneous regions, when compared to previously proposed methods.
Fast image recovery using variable splitting and constrained optimization	José BIOUCAS-DIAS Professeur Instituto Superior Técnico, Lisbonne	05/07/2010 14h30 salle Euler violet
Résumé (anglais) : In this talk, I will address a new fast algorithm for solving one of the standard formulations of image restoration and reconstruction which consists of an unconstrained optimization problem where the objective includes a convex data-fidelity term and a convex non-smooth regularizer. The formulation allows both wavelet-based (with orthogonal or frame-based representations) regularization or total-variation regularization. Our approach is based on a variable splitting to obtain an equivalent constrained optimization formulation, which is then addressed with an augmented Lagrangian method. The proposed algorithm is an instance of the so-called "alternating direction method of multipliers", for which convergence has been proved. Experiments on a set of image restoration and reconstruction benchmark problems show that the proposed algorithm is faster than the current state of the art methods.
Seismic fault detection using marked point processes	Barna KERESZTES Post-doctorant Université Technique de Cluj-Napoca, Roumanie et Université de Bordeaux 1	14/06/2010 14h30 salle Euler Violet
Résumé (anglais) : During this seminar a new method for seismic fault segmentation using stochastic processes will be presented. The model is a marked point process in which each fault is modeled by a curve. Three classical fault attributes are proposed to measure the likelihood of a fault. The computation of the likelihood can be based on one of the attributes as fault detector or on a composite function using the three measures. For fault detection in real 3D blocks we made several modifications to the original marked point process algorithm, to get a result which will be easier to interpret by the geologists. The results demonstrate the effectiveness of our approach. Using marked point process allows overcoming the drawbacks which are observed with the direct use of classical fault attributes.
Séminaire ANNULE	C. BHATTACHARYA Chercheur permanent Defence Institute of Advanced Technology, Pune, Inde	04/06/2010
Challenges in the analysis of light microscopy data	Peter HORVATH Chercheur Ecole polytechnique fédérale de Zurich	31/05/2010 14h30 salle Euler Violet
Résumé (anglais) : The recent evolution of imaging equipment and technology established light microscopy as one of the most important research tools for a number of biological/pharmaceutical research fields. Extraction of quantitative data from the acquired images (a basic requirement of modern biology) is however still a very challenging task with a variety of problems to be solved. In my talk I'll concentrate on two main research/application fields: RNAi (RNA interference) based high-content screening and high-end time-lapse (fluorescent) imaging. In the first case I will briefly introduce the basics of the technique and discuss the methods used for image analysis, statistics, bioinformatics, and quality control. I will present our latest results on virus/bacterial infection, cancer development and ribosome biogenesis. For the second part of my talk, I will concentrate on image segmentation and tracking problems related to fluorescent time-lapse (live cell) microscopy.
Full-waveform laser scanning: technology, processes, and applications	Clément MALLET Doctorant Institut Géographique National	03/05/2010 14h30 salle Coriolis
Résumé (anglais) : The new technology of airborne full-waveform laser scanning (lidar) systems has emerged in the last fifteen years, and has become popular the last five years. It permits to record the received signal for each transmitted laser pulse, the result is called a waveform. Such sample sequence represents the progress of the laser pulse as it interacts with the reflecting surfaces. Hence, FW lidar data yield more than a basic geometric representation of the Earth topography. Instead of clouds of individual 3D points, lidar devices provide connected 1D profiles of the 3D scene, which allows gaining further insight into the structure of the scene. Indeed, each signal consists of series of temporal modes, where each of them corresponds to the reflection from a unique object or a superposition of the signal of several elements. The advantage of off-line waveform processing is twofold: (1) Maximize the detection rate of relevant peaks within the waveforms. More points can be extracted in a more reliable and accurate way. (2) Decompose the waveforms by modeling each echo with a suitable parametric function. The echo shape can be retrieved, providing relevant features for subsequent segmentation and classification purposes. Waveform processing capabilities can therefore be extended by enhancing information extraction from the raw signals. The first issue is the waveform processing and modeling step. Two methods will be presented. On the one hand, the lidar waveforms can be considered as a Gaussian Mixture Model, based on simple physical assumptions. On the other hand, when such assumptions are not satisfactory, we aim to model each mode of the waveform with the best-fit analytical function of a given a set of parametric curves. A method based on a marked point process model that hypothesizes mixtures of various parametric functions will therefore be presented. Results over both urban and forested areas will be shown. The second issue deals with the integration of the parameters extracted from the modeling step into standard classification and surface reconstruction lidar algorithms. The relevance of full-waveform lidar features for classification of urban and mountainous areas will be presented using a Support Vector Machine classifier. Then, this relevance will be quantified against optical and traditional lidar data in urban areas with a Random Forest classifier. Finally, we will show how useful full-waveform features can be for Digital Terrain Model generation in forested areas, for both hydrological and archaeological purposes.
Reliability and accuracy in stereovision	Neus SABATER Post-Doctorante CMLA, École Normale Supérieure de Cachan	01/03/2010 14h30 salle Coriolis
Résumé (anglais) : This work is a contribution to stereovision written in the framework of the MISS (Mathematics for Stereoscopic Space Imaging) project launched by CNES in cooperation with several university laboratories in 2007. This project has the ambitious goal to model a stereo satellite, using two almost simultaneous views of the Earth with small baseline in urban areas. Its main goal is to get an automatic chain of urban reconstruction at high resolution from such pairs of views. The project faces fundamental problems that our work aims at solving. The first problem is the rejection of matches that could occur just by chance, particularly in shadows or occlusions, and the rejection of moving objects (vehicles, pedestrians, etc.). In this work we have proposed a method for rejecting false matches based on the "a contrario" methodology. The mathematical consistency of this rejection method will be shown and it will be validated on exact simulated pairs, on ground truths provided by CNES, and pairs of classical benchmark (Middlebury). The reliable accepted matches reach a 40% to 9% density in the tested pairs. The second issue is the accuracy. Indeed, the type of considered stereoscopy requires a very low baseline between the two views, which are visually almost identical. To get a proper relief, an extremely accurate shift must be estimated, and the noise level that allows this accuracy must be calibrated. In this work a subpixel disparity estimation method is proposed, which will be proved optimal by experimental and mathematical arguments. These results extend and improve the results obtained by the CNES method MARC.
Modélisation mathématique des textures : Oscillations, textons, ou motifs orientés ?	Jean-François AUJOL Chercheur CNRS LATP, CMI, Université de Provence	01/02/2010 14h30 salle Coriolis
Résumé (français) : Dans cet exposé, nous nous intéressons à la modélisation mathématique des textures. Nous proposons différentes approches, fréquentielles ou spatiales, pour décrire les textures. Notre étude sera illustrée de nombreux exemples en décomposition d'images en géométrie et texture, ainsi qu'en inpainting. Les résultats présentés sont le fruit de collaborations avec Vincent Duval (Telecom Paris Tech), Yann Gousseau (Telecom Paris Tech), Pierre Maurel (IRISA), et Gabriel Peyré (Ceremade).
Bayesian detection using the Posterior probability of the Likelihood Ratio. Application to exoplanet detection using direct imaging.	Isabelle SMITH Doctorante Universite de Nice Sophia-Antipolis, Département Astrophysique	25/01/2010 14h30 salle Coriolis
Résumé (anglais) : Extra-solar planet detection using direct imaging is very challenging: the luminosity contrast between the star and the exoplanet is very high and the exoplanet is very close to its parent star. To detect so low signals, high technology instruments like the VLT planet finder SPHERE are under development as well as sophisticated signal processing tools. Statistical detectors are expected to be robust with respect to the data and to have a sufficiently high Probability of good Detection (PD) for a given Probability of False Alarm (PFA). In order to calibrate common detectors, the *Posterior probability of the Likelihood Ratio** (PLR)* has been proposed in 1997 by Aitkin. In this presentation, the following analytical properties of the PLR will be presented for the simple versus composite hypothesis testing: - the computation of the PLR can be performed straightforwardly by sampling the posterior probability of the tested parameter - the support of the PLR is limited by the Generalized Likelihood Ratio Test - for a proper prior, the mean (and variance) of the PLR is equal to the (Fractional) Bayes Factor - for the test "Reject H1 if PLR > p" and for some invariant likelihood families and their uninformative priors, we have simply "PFA = 1-p" Finally, the PLR will be applied to exoplanet detection using direct imaging: a hierarchical model of the images will be presented as well as an illustration of the performance of the PLR used as a detector.