Laboratoire d'Informatique, Signaux et Systèmes de Sophia Antipolis

Lauréat catégorie recherche

DNA storage suffers from deletion, insertion, and substitution errors that may affect the sequence of nucleotides during synthesis, sequencing, and storage. Dealing with these errors is one of the main challenges in DNA storage since their presence can be detrimental to the reliability of the system. For instance, DNA sequencing often results in multiple erroneous reads of the data. Current approaches for recovering the data rely on multiple sequence alignment methods that are solely based on heuristics and do not scale well beyond small proof-of-principle experiments. Alternatively, reconstructing the data in question can be cast in the setting of trace reconstruction, which is a well-studied problem in computer science. In this talk, as a first step towards adapting trace reconstruction schemes to DNA storage, we present an efficient novel trace reconstruction scheme that provides robustness against deletions through coding-theoretic and algorithmic techniques.

These results will be presented during a seminar given on April 6th at 14h in the Conference Room (I3S-007) by Dr. Serge Kas Hanna, Senior Researcher at Aalto University.

Intelligent reconfigurable surfaces for wireless communications :
Reconfigurable intelligent surface (RIS) is a candidate technology for future wireless networks. It enables to shape the wireless environment to reach massive connectivity and enhanced data rate.
This talk is aimed for general audience, with a zoom link : https://univ-cotedazur.zoom.us/my/deneire

Guillaume Ducard, MCF HC, Dr. sc. ETH, will defend this Thursday his habilitation thesis (HDR) detailing his contributions to the research theme of "flight control, estimation, and fault diagnosis systems for unmanned aerial vehicles (UAVs)". The defense is open to the public and will take place at Auditorium F200 Forum Polytech in Sophia Antipolis.
Summary:
Unmanned aerial vehicles are spreading through our modern societies in countless new applications. However, for these UAVs to be routinely inserted in the civilian airspace, a key factor must be guaranteed: safety. This calls for flight controllers which are actuator- and sensor- fault tolerant. This research work presents the design of active fault-tolerant flight controllers. They are based on methods to actively monitor the health of actuators and sensors of a UAV. Once the fault diagnosis is done, the control laws are reconfigured. Three new fault detection and Isolation (FDI) techniques are presented:  1) the EMMAE method, 2) the SMAC-FDI algorithm, and 3) a FDI based on a single KF with hidden Markov models. Three different flying platforms are considered: 1) fixed-wing aircraft, 2) multi-rotor helicopters, and 3) hybrid VTOL UAVs. Nonlinear flight controllers are developed for each platform. In addition, for multi-rotor helicopters, a new attitude estimator and a new control allocation (CA) algorithm are developed. The new CA is based on a weighted pseudo-inverse matrix method, called ”WCA”. WCA outperforms all current control allocation techniques based on classical pseudo-inverse methods. For hybrid or convertible UAVs, three different flight controllers are designed: 1) a Fused-PID control approach, 2) a MPC flight controller, and 3) a neural-network controller. They are all successfully flight tested in real experiments on a tilt-rotor vertical take-off and landing (VTOL) platform. Finally, new developments around neural-network-inspired controllers are presented and validated on academic examples, such as an inverted pendulum or a crane for payload transport.

How structural connectivity (SC) gives rise to functional connectivity (FC) is not fully understood. Here we mathematically derive a simple relationship between SC measured from diffusion tensor imaging, and FC from resting state fMRI. We establish that SC and FC are related via (structural) Laplacian spectra, whereby FC and SC share eigenvectors and their eigenvalues are exponentially related. This gives, for the first time, a simple and analytical relationship between the graph spectra of structural and functional networks. Laplacian eigenvectors are shown to be good predictors of functional eigenvectors and networks based on independent component analysis of functional time series. A small number of Laplacian eigenmodes are shown to be sufficient to reconstruct FC matrices, serving as basis functions. This approach is fast, and requires no time-consuming simulations. It was tested on two empirical SC/FC datasets, and was found to significantly outperform generative model simulations of coupled neural masses.

https://univ-cotedazur.zoom.us/j/88571081975
demandez le code ! (fpayan@i3s.unice.fr)
 

farras.abdelnour@gmail.com
http://sacklerinstitute.academia.edu/FarrasAbdelnour
http://www.abdelnour.photos
 

This presentation will outline contributions to the research theme of direct visual localisation and mapping carried out at the national scientific research centre (CNRS). The central theme of this research began with the development of a first real-time large-scale direct visual SLAM approach in 2007 that paved the way for subsequent innovation and results. A historical and bibliographic overview of direct SLAM is provided in the introduction followed by the interest of this approach for various real-world applications including autonomous driving, intervention scenarios, space robotics, humanoid manufacturing, drones and augmented reality. The contributions have been categorised into three main groups: sensor modelling, localisation, and lastly mapping. An overview of past research projects that have been significant to direct SLAM including the industrial transfer and creation of the start-up PIXMAP will be presented. A succinct overview of fundamentals and definitions will be first provided. Visual sensor models will be introduced ranging from basic cameras to generalised multi-view camera systems with rolling shutter and motion blur effects. Localisation methods will be covered next including pose estimation from both colour and depth. Extension to large scales, dynamic localisation and high-speed approaches will also be presented. The following focuses on mapping approaches including light fields, key-frame graphs, high-dynamic range mapping and super-resolution. To finish future research goals and an overview of projects that are currently commencing will be presented.

HDR committee:
- Reviewers:
Tom DRUMMOND, Professor, Melbourne University, Australia
Simon LACROIX, CNRS Research Director, LAAS Toulouse
José NEIRA, Professor, University of Zaragoza, Spain

- Examiners:
Ezio MALIS, Research Director, INRIA Sophia-Antipolis Méditerannée
Eric MARCHAND, Professor, University of Rennes 1
Tarek HAMEL, Professor, University Côte d'Azur

.

David will present the overall context of the new concepts emerging in term of risk assessment of chemicals and Fabrice will present his project on the use of cell painting data for toxicity prediction.

Reducing acquisition time is a major challenge in high-resolution MRI that has been successfully addressed by Compressed Sensing (CS) theory. While the scan time has been massively accelerated, the complexity of image recovery algorithms has strongly increased, resulting in slower reconstruction processes. In this work we propose an online approach to shorten image reconstruction times in the CS setting. We leverage the segmented acquisition in multiple shots of k-space data to interleave the MR acquisition and image reconstruction steps. This approach is particularly appealing for 2D high-resolution T2* -weighted anatomical imaging. During the scan, acquired shots are stacked together to form mini-batches and image reconstruction may start from incomplete data. We demonstrate the interest and time savings of this online  image reconstruction framework for Cartesian and non-Cartesian sampling strategies combined with a single receiver coil. Next, we further extend this formalism to address the more challenging case of  multi-receiver phased array acquisition. In this setting, calibrationless image reconstruction is adopted to remain compatible with the timing constraints of online delivery. Our results on ex-vivo 2D T2* -weighted brain images show that high-quality MR images are recovered by the end of acquisition in both acquisition setups. 

Depuis le travail de Y. Nesterov (1984), et surtout l’algorithme FISTA de Beck et Teboulle (2008) ou l’algorithme rapide de Weiss, Blanc-Féraud, Aubert (2008) , il est reconnu qu’utiliser un algorithme de gradient inertiel est beaucoup plus efficace pour minimiser une fonctionnelle convexe qu’un simple algorithme de descente de gradient. Nous verrons qu’en fait l’utilité de l’inertie dépend très fortement de la géométrie au voisinage du minimiseur de la fonctionnelle, et qu’il n’est pas toujours préférable d’utiliser un terme inertiel. Ces résultats ont des conséquences directes en traitement d’images et en deep learning. Il s’agit de travaux en collaboration avec Vassilis Apidopoulos (IMB), Charles (IMT, INSA) Dossal, et Aude Rondepierre (IMT, INSA).

The link between genetic regulation and morphogenesis of multicellular organisms is a fundamental problem in developmental biology. We address this question in the shoot apical meristem (SAM) of vascular plants, which plays a central role in morphogenesis by generating plants above ground organs through complex changes in growth rates and directions of its individual cells. Although the dynamics and structure of gene regulatory networks in the SAM have been extensively studied, little is known about the spatiotemporal correlation between gene expression patterns and growth at cellular and tissue scale. This is primarily due to the lack of comprehensive quantication of organ wide cell properties over time with regards to underlying molecular networks. By focusing on morphogenesis at the Arabidopsis thaliana in fluorescence meristem, using confocal time-lapse live imaging, we computed cell lineages and cell properties such as volumes, anisotropy, surface area as well as growth and division patterns. These results were integrated in a three dimensional, dynamic template or atlas of a developing primordium. To link morphogenesis to the underlying gene networks, we also integrated the expression patterns of genes that control flower development into this atlas. Thus the expression patterns of 27 genes involved in organ identity, growth control and organ polarity were recorded in the template. A detailed, quantitative correlation analysis between gene expression patterns and growth patterns allowed us to propose speficic hypotheses for the combinatorial action of regulatory genes.

Nombre de mesures expérimentales sont altérées par des fluctuations stochastiques ou systémiques, souvent inhérentes aux protocoles expérimentaux et aux méthodes d'acquisition. Dans ce contexte, nous présentons BEADS, une méthode permettant de séparer les sources d'une observation composée de manière ternaire : un signal d'intérêt (plutôt parcimonieux) formé d'une superposition de pics, une tendance, dérive oul igne de base relativement de basse fréquence, et du bruit. Ce modèle est relativement courant pour les signaux issus de l'analyse physico-chimique, qui forment le coeur du projet AMIDEX BIFROST (Blind Identification, Filtering & Restoration On Spectral Techniques).

En s'appuyant sur différents a priori morphologiques sur les signaux (ainsi que sur leurs dérivées) comme la linéarité, la positivité, ou encore la parcimonie, ce problème est formulé comme la minimisation d'une fonction comportant un terme quadratique de fidélité aux données, d'une pénalité régularisée  de type norme $\ell_1$ asymétrique promotrice de positivité,  ainsi qu'une contrainte de parcimonie. Nous nous intéressons ici au cas de signaux physico-chimiques, et particulièrement à la chromatographie, où les chromatogrammes 1D et 2D sont bruités et présentent de fortes dérives de la ligne de base, biaisant ainsi les informations qui peuvent en être extraites. Les performances sont évaluées sur des données simulées et réelles.
Cette modélisation  étant assez générique, BEADS a été appliquée à d'autres types de signaux chimiques, en astronomie, dans le domaine biomédical ou le monitoring.
Nous évoquerons enfin les travaux en cours sur l'estimation des paramètres, le recalage élastique pour des données chromatographiques 2D, et des perpectives déconvolution aveugle avec une pénalité en rapport régularisé de normes.

Références
**"Chromatogram baseline estimation and denoising using sparsity (BEADS)", Xiaoran Ning, Ivan W. Selesnick, Laurent Duval, Chemometrics and Intelligent Laboratory Systems, December 2014
http://dx.doi.org/10.1016/j.chemolab.2014.09.014

**BEADS: Baseline Estimation And Denoising with Sparsity
http://www.laurent-duval.eu/siva-beads-baseline-background-removal-filtering-sparsity.html

Abstract
 
The proliferation of autonomous robots evidence forthcoming environments where multiple autonomous systems (MAS) will be interacting with each other, as well as with human beings, to perform complex tasks at a level never imagined before. Conventional methods for improving MAS performance address very specific challenges, but not general problems. Learning-based controllers offer adaptability and robustness against uncertainties, however, the computational complexity of these solutions is often not practically feasible. These drawbacks limit the applicability and penalize the performance of current MAS control methods. Recently, cognitive scientists advocate that “a single occurrence of an emotionally significant situation is remembered far more vividly and for a longer period than a task, which is repeated frequently”. This highlights that emotional processing is able to develop an effect that sustained sensory input is not able to achieve. In this talk, we present conventional and adaptive distributed consensus algorithms for MAS. Next, a descriptive and a mathematical model of emotion processing in the mammalian brain is introduced, which is then modified to develop a hierarchical feedback control for MAS. Preliminary results show how the basic features of the emotional learning system in combination with the MAS controller can help to effectively control a group of robots in real-time, in presence of system uncertainties.
 
 
Speaker’s Bio
 
Luis Rodolfo Garcia Carrillo was born in Durango, Mexico in 1980. He received the Licenciatura in Electronic Engineering in 2003, and the M.Sc. in Electrical Engineering in 2007, both from the Instituto Tecnologico de La Laguna, in Coahuila, Mexico. He received his Ph.D. in Control Systems from the University of Technology of Compiegne, France, in 2011, where he was advised by Professor Rogelio Lozano. From 2012 to 2013, he was a postdoctoral researcher at the Center for Control, Dynamical Systems and Computation at the University of California, Santa Barbara, where he was working with Professor Joao Hespanha. He currently holds an Assistant Professor position with the Department of Electrical Engineering at Texas A&M University – Corpus Christi. His current research interests include multi-agent control systems, intelligent controllers, and vision-based control.

L'apprentissage profond ("Deep Learning") est actuellement très en vogue. Cependant, ses fondements théoriques restent complexes et mal expliqués. Les théories en traitement du signal et des images pourraient sans doute contribuer à améliorer l'explicabilité de l'apprentissage profond. Récemment, Stéphane Mallat et plusieurs de ses collègues ont introduit la notion de "Deep Scattering Network" qui s'appuie sur l'utilisation d'ondelettes pour éclairer certains aspects théoriques d'une architecture profonde. Le but de cet exposé est de fournir une brève introduction au "Deep Scattering Network" et d'ouvrir une discussion sur les liens possibles entre le traitement du signal et l'apprentissage profond.

Entropy has been extensively characterized when analyzing Heart Rate Variability (HRV) series, and, generally, time series in diverse scientific fields. Many metrics were proposed in the last 30 years, but, in practice, a common critical point is the selection of the parameters on which these techniques rely on. In this talk, a new definition of entropy, aiming to reduce the significance of this selection, is presented. Bubble Entropy is based on permutation entropy, where the vectors in the embedding space are ranked. Using bubble sort we count instead the number of swaps performed for each vector. As a consequence, the definition proposed is almost free of parameters. Experimental results showed that bubble entropy presents remarkable stability and exhibits increased descriptive and discriminating power compared to other definitions.

CV : https://homes.di.unimi.it/sassi/CV_eng.pdf

Roberto Sassi is Associate Professor of Computer Science in the Department of Computer Science at the University of Milan, Italy, where he teaches courses on digital signal processing and a master class on biomedical signal processing. He is a member of the doctoral school in Computer Science and has supervised 5 Ph.D. students. He is an IEEE senior member since 2012 and an active member of the e-Cardiology working group of the European Society of Cardiology, for which he participated in coordinating the development of the recent technical consensus document on heart rate variability. The research activity of Dr. Sassi has been mainly focused in the fields of biomedical signal processing and applied mathematics. The methodological techniques have been exploited also to address related challenges in image processing and computer science, mainly biometrics. The results have been published in more than 100 publications in international journals and conferences. 

Présentation d'Oscar

Dans cet expose, je présenterai une vue globale de mes travaux de 
recherches. Ils ont été motives par des problématiques de traitement 
d'images. Récemment, mon champ d’intérêt s’est agrandi à l’analyse de 
données en grandes dimensions et aux problèmes d’apprentissage.

Le problème de restauration d'images dégradées par des flous variables 
connaît un attrait croissant et touche plusieurs domaines tels que 
l'astronomie, la vision par ordinateur et la microscopie à feuille de 
lumière où les images peuvent être constituées de milliards de voxels. 
Les flous variables peuvent être modélisés par des opérateurs intégraux 
linéaires. Une fois discrétisés pour être appliqués sur des images de N 
pixels, ces opérateurs peuvent être vus comme des matrices gigantesques 
de taille N x N. Pour les applications visées, les matrices contiennent 
10^(18) coefficients. On voit apparaître ici les difficultés liées à ce 
problème de restauration des images qui sont i) le stockage de ce grand 
volume de données, ii) les coûts de calculs prohibitifs des produits 
matrice-vecteur. On dit que ce problème souffre du fléau de la 
dimension. De plus, dans beaucoup d'applications, l'opérateur de flou 
n'est pas ou que partiellement connu. Il faut donc simultanément estimer 
et approcher ces opérateurs intégraux.

Je présenterai de nouveaux modèles et méthodes numériques avec des 
garanties théoriques fortes permettant de traiter ces difficultés.

Présentation de Signet

Présentation de Morpheme

We present a framework to design an orthogonal wavelet with compact support and vanishing moments, tuned to a given application in a data-driven way.
This is achieved by optimizing a criterion, such that a prototype signal, which is characteristic for the application, becomes sparse in the wavelet domain.
This approach is beneficial for compression and detection purposes.
Starting from a filter bank approach with lossless polyphase matrices, a parameterization is developed for which orthogonality and compact support are built in, and in terms of which we can express the vanishing moment conditions conveniently. For low order filters the vanishing moment conditions can be built in as well, for high order filters one should resort to constrained optimization. The approach is developed for critically sampled wavelet transforms as well as for the stationary wavelet transform. A few examples illustrate the wavelets generated by these methods.

Présentation de Mediacoding

L'objectif de ce travail est d'être capable de détecter des nano-particules dans des images industrielles et de définir des indicateurs statistiques, et des procédures d'estimation associées, permettant de quantifier la répartition de ces nano-particules (répartition spatiale, agrégation ou pas....).

Nous avons effectué une phase de preprocessing des données. Les images industrielles dont nous disposons étant issues de mesures par microscopie à force atomique, nous devons corriger des dérives de l'appareil de mesure.

Nous cherchons ensuite à estimer des caractéristiques de nos nanos objets en combinant ces mesure AFM avec une autre technique d’imagerie (MEB). J’expliquerai notre premiers résultats qui sont basés sur des techniques d'agrégation d'estimateurs

Présentation de Design

Présentation des activités de l'équipe SIGNAL

En classification traditionnelle, chaque échantillon à classer se voit attribuer une classe unique, contexte dit mono-label. La classification multi-labels représente une extension où un échantillon à classer peut être associé à plusieurs classes simultanément.

Dans le domaine médical, l'application de la classification multi-labels s'avère particulièrement adaptée puisque les patients peuvent être atteints de plusieurs pathologies simultanément. Or, l'utilisation d'une méthode mono-label par pathologie ne permet pas de prendre en compte les corrélations potentielles entre pathologies. Nous verrons les deux grandes familles de méthodes multi-labels (les méthodes d'adaptation et par transformation) ainsi qu'une application aux données MAPA (Mesure Ambulatoire de la Pression Artérielle).

Title: Introduction to multi-label classification: application to medical data

Abstract:
In traditional classification, each sample to be classified is assigned a single class, also called mono-label classification. The multi-label classification represents an extension where a sample can be associated with several classes simultaneously.

In the medical field, the application of multi-label classification is particularly adapted since patients may be suffering from several pathologies simultaneously. Indeed, the use of one mono-label method per pathology does not allow to take into account the potential correlations between pathologies. We will see the two main families of multi-label methods (adaptation and transformation methods) as well as an application to ABPM data (Ambulatory Blood Pressure Monitoring).

Many problems in signal and image processing can be formulated as the minimization of a cost functional with additional constraints related to prior knowledge. The most simple such functional being ||Ax-d|| (where d is some measurement) possibly subject to a linear constraint on x, which has an exact solution provided by the least squares method. In other applications a sparsity constraint on x is more relevant, leading to a non-convex functional f(x)=||Ax-d||+"the amount of nonzero elements in x". The corresponding l0-l2 minimization problem is hard, non-convex and usually has a multitude of local minima. Another problem in the same category is that of finding a low rank matrix satisfying certain additional restrictions. 

In my talk I will present an overview of such problems, provide theory for the convex envelope of the functional f and show how the convex envelope can be used to find approximate solutions to the problem, as well as examples of how the convex envelope can help us understand the performance (and prove convergence) of algorithms designed for convex functionals, applied to non-convex ones. 

La  reconstruction  géométrique  d'une  scène  basée  sur  des  données  3D  est  importante  car  elle  possède  de  nombreuses applications. Dans un premier temps, il faut détecter l’ensemble des formes primitives (par exemple cylindres, sphères, cônes,  etc.)  qui  composent  la  scène  3D  avant  d’ensuite  tenter  d’établir  des  relations  entre  elles.  Les  formes  primitives détectées automatiquement par les méthodes existantes sont souvent assez limitées.

Je présente une méthode tirant profit de la symétrie des surfaces de révolution, afin de reconstruire la géométrie  de  la  surface  de  révolution  de  manière  précise  et  efficace.  J'illustre  la  pertinence  de  la  méthode  en reconstruisant des surfaces de révolution en temps réel dans un environnement issu d’un SLAM dense. 

La  Réalité  Augmentée  vise  à  afficher  des  informations  numériques  virtuelles  sur  des  images  réelles.  Le  recalage  est important, puisqu’il permet d'aligner correctement les objets virtuels dans le monde réel. Contrairement au tracking qui recale  en  utilisant  les  informations  de  l’image  précédente,  la  localisation  à  grande  échelle  (wide  baseline  localization)
calcule la solution en utilisant uniquement les informations présentes dans l’image courante. Il permet ainsi de trouver des solutions initiales au problème de recalage (initialisation) et, n’est pas sujet aux problèmes de « perte de tracking ». Le problème  du  recalage  en  RA  est  relativement  bien  étudié  dans  la  littérature,  mais  les  méthodes  existantes  fonctionnent principalement  lorsque  la  scène  augmentée  présente  des  textures.  Pourtant,  pour  le  recalage  avec  les  objets  peu  ou  pas texturés, il est possible d’utiliser leurs informations géométriques qui représentent des caractéristiques plus intrinsèques que les textures. Ce séminaire s’attache au problème de recalage basé sur des informations géométriques, et plus précisément sur les  points.  Je  présente  LGC,  une  nouvelle  méthode  de  recalage  de  points  robustes  et  rapides.  Il  peut  mettre  en correspondance des ensembles de motifs de points 2D ou 3D subissant une transformation dont le type est connu. LGC présente un comportement linéaire en fonction du nombre de points, ce qui permet un tracking en teps-réel. Je montrerai un démo d'augmentation de croquis d'ingénieries.

Because of atmospheric turbulence, images of objects in outer space acquired via ground-based telescopes are usually blurry. One way to estimate the blurring kernel or point spread function (PSF) is to make use of the aberration of wavefront received at the telescope, i.e., the phase. However only the low-resolution wavefront gradients can be collected by wavefront sensors. In this talk, I will discuss how to use regularization methods to reconstruct high-resolution phase gradients and then use them to recover the phase and the PSF in high accuracy.

Abstract : Proton therapy kills cancer cells by delivering proton beams to the tumour. Opposite to conventional radiation therapy, proton beams deposit their maximum energy within a precisely defined range, known as the Bragg peak. This is due to the stopping power of the atoms in the body of the patient. To proceed the treatment, a plan is draw on a 3D image (a CT scan) of the patient. The Gross Tumour Volum is segmented and the Organs At Risk (OAR) are also identified to be protected during the treatment.  The plan allows to maximize the dosis on the tumour and minimize the dosis on the OAR. The patient is treated in the room on the basis of the 3D scan taken before the treatment.
Our research has led to the development of an imaging system (a cone beam CT) which is in the treatment room and which allows to adapt the treatment by using different image processing tools, among which deformable coregistration and  contours tracking are essential. This allow to implement adaptive protontherapy, including on mobile tumours (e.g. lung tumours) and reduce the range uncertainties.

Biography : Benoit Macq is Professor at Université catholique de Louvain (UCL). His main research topics are image and video compression, image watermarking, immersive communications, visualization and co-registration for medical imaging.
Benoit Macq is Fellow of the IEEE and has been general chair of IEEE ICIP2011.
Benoit Macq has been Vice-President of UCL from 2009 to 2014 and has founded the Louvain Technology Transfer Office.
Benoit Macq is co-founder of 10 start-ups from his research team.
He is Member of the Royal Academy of Science of Belgium.

Abstract: Principal component analysis (PCA) based on L1-norm maximization is an emerging technique that has drawn growing interest in the signal processing and machine learning research communities, especially due to its robustness to outliers. In this talk we will explore some links between L1-norm PCA and ICA. Specifically, we will show that L1-norm PCA can perform independent component analysis (ICA) under the whitening assumption. However, when the source probability distributions fulfil certain conditions, the L1-norm criterion needs to be minimized rather than maximized, which can be accomplished by simple modifications on existing optimal algorithms for L1- PCA with global convergence. If the sources have symmetric distributions, we show in addition that L1-PCA is linked to kurtosis optimization. A number of numerical experiments illustrate the theoretical results and analyze the comparative performance of different algorithms for ICA via L1-PCA.

Reference:
R. Martín-Clemente, V. Zarzoso, "On the Link between L1-PCA and ICA", IEEE Transactions on Pattern Analysis and Machine Intelligence, 38, 2016.
 
Short bio: Rubén Martín-Clemente received the M.Eng. Degree in telecommunications engineering and the PhD degree with highest distinction in telecommunications engineering from the University of Seville, Spain, in 1996 and 2000 respectively. Currently, he is associate professor at the Department of Signal Processing and Communications of the University of Seville, Spain. He has been a visiting researcher at the University of Regensburg, Germany. Among other areas, his research interests include multivariate data analysis with emphasis on independent component analysis and its application to biomedical problems. He has authored or co-authored numerous publications on these topics. Dr. Martín-Clemente has served as Program Committee Member for several international conferences and was a Program Committee Chair of the 5th International Conference on Independent Component Analysis and Blind Signal Separation in 2004.

Several respiratory, cardiovascular and metabolic disorders are diagnosed and monitored by respiratory measurements such as respiratory rate, tidal volume and breathing pattern. The methods to retrieve these parameters usually require some obtrusion of the air breathed by the patients with is uncomfortable and limited to hospital environments. New respiratory measurement techniques are being developed to overcome these limitations, and one emerging approach involves the use of accelerometers to analyze the thoracic cage movement and estimate the respiratory parameters. We discuss these strategies and also introduce a new system for measure multiple accelerometer signal simultaneously and, using blind source separation, estimate these respiratory parameters.

Biography: Ailton L. D. Siqueira Jr. is an Electrical Engineer, and has received the M. Sc. (2007) and Ph. D. (2013) degrees in biomedical engineering from the Federal University of Uberlândia (Brazil). He is a full time professor on Federal Institute of Triângulo Mineiro since 2010, and currently is a Post-Doctoral Researcher with the I3S SIS team. His research interests include electrical engineering, biomedical engineering, computer systems, digital signal processing, electromyography and medical instrumentation.

Non obstructive acquisition of respiratory waveforms has applications in several research and clinical scenarios. This presentation will introduce two research fields that can benefit of such system: computerized lung sound analysis and autonomic cardiovascular regulation modeling.

Biography: Raimes Moraes received the B.Sc. degree in electrical engineering from the Federal University of Uberlândia (Brazil) in 1988, the M.Sc. degree in electrical engineering from the State University of Campinas (Brazil), in 1991, and the Ph.D. degree in medical physics from the University of Leicester, (U.K.), in 1995. He is Professor of the Electrical Engineering Department, Federal University of Santa Catarina (Brazil).  His research interests include medical instrumentation and discrete-time signal processing applied to biomedical signals.

Abstract: Quadrotor aerial vehicles are one of the most flexible and adaptable platforms for aerial robotics research. The impact of the quadrotor in the field of robotics research can be seen as similar to that of the Puma robotic arm in the early years of robotic manipulators, and the unicycle wheeled robot, in a similar period in mobile robotics. Existing research aerial robotics has brought the field to a point where the basics of control and estimation are reasonably well established for quadrotor vehicles although contributions to the basic control design are still being made. This talk provides a high level introduction to the key aspects of quadrotor control as practiced by the majority of laboratories around the world.

Biography:
Robert Mahony is a Professor in the Research School of Engineering  at the Australian National University.  He received his BSc in 1989 (applied mathematics and geology) and his PhD in 1995 (systems  engineering) both from the Australian National University.  He worked firstly as a marine seismic geophysicist and an  industrial research scientist before completing a postdoctoral  fellowship in France at the Universite de Technologie de Compiegene and a Logan Fellowship at Monash University in Australia.  He has  held his post at ANU since 2001.  His research interests are in  non-linear systems theory and optimization with applications in  robotics, geometric optimisation techniques and computer vision. 

This talk will cover two applications related to digitization of the real world: one concerning cultural heritage and another on example-based texture synthesis.

I) Treatment of 3D models with acquired radiance.
Vision and computer graphics communities have built methods for digitizing, processing and rendering 3D objects. There is an increasing demand coming from cultural communities for these technologies, especially for archiving, remote studying and restoring cultural artefacts like statues, buildings or caves. Besides digitizing geometry, there can be a demand for recovering the photometry with more or less complexity: simple textures (2D), light fields (4D), SV-BRDF (6D), etc. This talk will cover several contributions that can be situated in the pipeline for constructing and treating surface light fields represented by hemispherical radiance functions attached to a surface. First, we will tackle robustness issues in the aspect reconstruction from photographic data resulting from real-world on-site acquisitions [VSLD13]. And secondly we will present a simplification technique for this data, which locally minimizes the loss of both geometric and photometric detail [VSKLD15].

II) Example-based real-time texturing.
Generating textures in real-time is a must in videogames or other real-time animated worlds. But to create textures quickly, they need to be generated at runtime from compact data stored on the GPU. In this talk we will explain why, and overview two contributions that improve over the state of the art in on-the-fly texture synthesis. The first [VSGLD13] is a new tiling algorithm that augments the amount of variety that can be generated with a single repetitive tile. This is achieved by stochastic exchanges of part of its content in a seamless way. Above that, a multi-scale transition mechanism is added to generate visual detail. The second [GSVDG14] will propose a new noise model for procedural texturing, which are textures stored as compact functions. This new function has been built so that it can be tuned to reproduce an input exemplar containing structured elements in a procedural way, which was unseen before.

REFERENCES
[VSLD13] Robust Fitting on Poorly Sampled Data for Surface Light Field Rendering and Image Relighting ; Computer Graphics Forum, vol. 32, issue 6.
[VSGLD13] On-the-Fly Multi-Scale Infinite Texturing from Example ; Proceedings of "ACM SIGGRAPH Asia 2013" conference ; Transactions on Graphics, vol. 31, issue 6.
[GSVDG14] Local random-phase noise for procedural texturing ; Proceedings of "ACM SIGGRAPH Asia 2014" conference ; Transactions on Graphics, vol. 32, issue 6.
[VSKLD15] Simplification of Meshes with Digitized Radiance ; Proceedings of the "Computer Graphics International 2015" conference ; The Visual Journal, vol. 21, issue 6-8.

 2ème exposé : mercredi 16 septembre, 10h00.

Résumé : This study was motivated by the biomechanical analysis of human movement, which usually consists of extensive reports with several graphs that are difficult to be interpreted by physicians or athletic trainers. Additionally, in many times, signal analysis is reduced to few parameters extracted from specific points of these signals, whose do not reflect or take into account the complexity of the involved physiologic phenomenon. The principal component analysis (PCA) has enabled the quantitative evaluation of signals by reducing large volumes of data to a small set of parameters, which take into account all the variance presented by the original data. Such parameters have been very useful for the quantitative evaluation of abnormalities, particularly allowing the objective comparison of clinical interventions or physical training strategies. More recently, the analysis of the loading factors of the eigenvectors obtained by PCA are representing an objective method to determine in which points of the signal a given subject shows abnormal differences, when compared to a control group. This brief presentation will include some examples of PCA application: (1) The gait analysis of subjects with Parkinson's syndrome for the objective comparison between treatments with drugs (usually levodopa) and deep brain stimulation of the subthalamic nucleus; (2) Assessment of physiotherapy effects in subjects after unilateral fracture of the lower limb; (3) Development of scores for normal gait and their application in functional evaluation of patients after reconstruction of the anterior cruciate ligament; and (4) It is finalized with an applications in ECG analysis.

 1er exposé : lundi 14 septembre, 14h00.

Résumé : This talk comprehends an overview of the Biomedical Engineering Program (PEB) at the Universidade Federal do Rio de Janeiro, including few details of our master and doctoral programs. It follows with a short presentation of our research areas: (1) Biomedical Instrumentation; (2) Health Systems Engineering; (3) Ultrasound in Medicine; (4) Pulmonary Engineering; (5) Biomechanics; and (6) Biomedical Signal Processing. For each area, details of some projects are given.

8h30-­‐9h Accueil café
Exposés :
• Présentation journée
• Projet MediaCoding, Marc Antonini
• « Contributions for image retrieval in the Wavelet Transform domain », Amani Chaker
• Projet Morpheme, Xavier Descombes
• « Modelling and characterizing axon growth from in vivo data », Agustina Razetti

10h45-­‐11h05 P’tite pause
Exposés :
• Projet Signal, Olivier Meste
• « Structured tensor decompositions and multidimensional compressed sensing », Henrique Goulard
• Projet Signet, Guillaume Urvoy-­‐Keller
• « Next Generation SDN-­‐based Virtualized Datacenters », Myriana Rifail

12h30-­‐14h Buffet
Exposés :
• « Bien vivre au quotidien avec l'équipe administrative », Marie-­‐Pierre Combeau
• Projet Design, Joao Rendas
• « Modélisation et prédiction du risque d'accident de décompression en plongée hyperbare », Asya Metelkina

15h20-­‐15h40 P’tite pause
Réunion des permanents
17h30 fin de la Journé

Making a judgment about the semantic category of a visual scene, such as whether it contains an animal, is typically assumed to involve high-level associative brain areas. Previous explanations require progressively analyzing the scene hierarchically at increasing levels of abstraction, from edge extraction to mid-level object recognition and then object categorization. Here we show that the statistics of edge co-occurrences alone are sufficient to perform a rough yet robust (translation, scale, and rotation invariant) scene categorization. We first extracted the edges from images using a scale-space analysis coupled with a sparse coding algorithm. We then computed the "association field" for different categories (natural, man-made, or containing an animal) by computing the statistics of edge co-occurrences. These differed strongly, with animal images having more curved configurations. We show that this geometry alone is sufficient for categorization, and that the pattern of errors made by humans is consistent with this procedure. Because these statistics could be measured as early as the primary visual cortex, the results challenge widely held assumptions about the flow of computations in the visual system. The results also suggest new algorithms for image classification and signal processing that exploit correlations between low-level structure and the underlying semantic category.

L’objectif du colloque est de mettre en avant des travaux en cours sur les éléments qui composent la chaîne de numérisation d’objets 3D, mais également de faire émerger de futurs axes de recherche dans les domaines de l’acquisition, de la modélisation, de l'analyse, et du traitement de la géométrie 3D pour le stockage, la manipulation, la visualisation etc.

Canonical Polyadic Decomposition (CPD), also known as PARAFAC, is a useful tool for tensor factorization. It has found application in several domains including signal processing and data mining. With the deluge of data faced in our societies, large-scale matrix and tensor factorizations become a crucial  issue. Few works have been devoted to large-scale tensor factorizations. In this paper, we introduce a fully distributed method to compute the CPD of a large-scale data tensor across a network of machines with limited computation resources. The proposed approach is based on collaboration between the machines in the network across the three modes of the data tensor. Such a multi-modal collaboration allows an essentially unique reconstruction of the factor matrices in an efficient way. We provide an analysis of the computation and communication cost of the proposed scheme and address the problem of minimizing communication costs while maximizing the use of available computation resources.