Gérard Giraudon
Inria Sophia Antipolis - Méditerranée (Research Centre Director)

Jan Koenderink
The Flemish Academic Centre for Science and the Arts (VLAC), Brussels, Belgium and Delft University of Technology, Delft, The Netherlands
"Perception as a User Interface"
Abstract:
Perception in the sense of “optical awareness” differs in kind from machine vision in that evolution does not drive systems to a veridical representation of the physical world, but to the construction of efficacious interfaces. Such interfaces determine idiosyncratic realities, awareness being a nexus of meaning. This, again, differs from machine vision, which is of a syntactical (meaningless) nature, meaning being imposed by a user or designer. In this sense machine vision emulates biological “vision as optically guided behavior”, which represents the automatic (not conscious, or “zombie”) functioning of agents. I will show a number of perhaps surprisingly non-veridical aspects of optical awareness, and will sketch ways to construct formal accounts of optical awareness in human observers.

Eric Grimson
Massachusetts Institute of Technology (USA)
"Computer Vision in Surgery and Disease Analysis"
Abstract:
Over the past few decades, computer vision techniques have played an increasing role in the practice of medicine. Segmentation methods that incorporate shape, appearance and atlas information provide detailed 3D reconstructions of patient-specific anatomy. Flexible registration methods align segmented models with patients for surgical planning and intra-operative guidance. Related registration methods align populations of images to support isolation of disease specific changes, as well as temporal evolution of those changes. We will review recent advances in these and related areas, highlighting the impact of computer vision methods in emerging medical practice.

Mike Brady
University of Oxford (UK)
"Some recent developments in Cancer Image Analysis using PET"
Abstract:
Positron Emission Tomography (PET) provides metabolic and potentially quantitative information. Its main clinical applications are currently in oncology, for example by enabling the detection of distal metastatic disease. However, despite increasing interest in molecular imaging, PET image analysis remains in its infancy. There are several inter-related challenges that need to be addressed in further developing PET for clinical and preclinical use, and several of them will be discussed: the reconstruction of time activity curves (of the uptake of the radioligand) at each voxel; the detection and repair of sudden "twitch" motions which normally cause (expensive) PET acquisitions to be discarded; the development of a biomathematical model to predict the utility of a proposed radioligand before it is injected into any animal; image fusion (increasingly PET-MRI); and the modelling of specific aspects of cancers, such as hypoxia, which refers to the remarkable ability of a tumour to grow rapidly even when it is starved of oxygen. An underlying theme of the talk will be to draw attention to parallels with the superb work done by Olivier Faugeras and colleagues on mathematical modelling of image formation and analysis.

Paul Bresloff
University of Oxford (UK)
"Geometric hallucinations and the functional architecture of primary visual cortex"
Abstract:
Geometric visual hallucinations are seen by many observers after taking hallucinogens, on viewing bright flickering lights, on waking up or falling asleep, in “near death” experiences, and in many other syndromes. We present a dynamical theory of their origin in primary visual cortex (area V1), based on the assumption that the form of the eye-brain map and the network architecture of V1 determine their geometry. The dynamics of V1 is described in terms of a continuum neural field model in which V1 is treated as a continuous two-dimensional sheet of neural tissue. Based on anatomical evidence, we take the network architecture to exhibit certain symmetries rendering it invariant under a novel action of the Euclidean group, composed of rigid body reflections, rotations and translations in the plane. Using this symmetry, we show how the common types of geometric hallucinations arise dynamically through the spontaneous formation of neural activity patterns within the cortical network. Our results thus suggest that the cortical mechanisms that generate geometric visual hallucinations are closely related to those used to process edges and contours in normal vision.

Yves Fregnac
Unité de Neuroscience, Information et Complexité, CNRS-UNIC (France)
"Functional polymorphy of visual cortical dynamics: from "crystal" to "smoke" L'irrégularité temporelle de l'activité spontanée est une caractéristique des réseaux sensoriels corticaux. Un problème encore non résolu est de comprendre dans quelle mesure cette activité apparentée à un “bruit” interfère avec la transmission du signal reçu de l'environnement et dégrade les capacités calculatoires du réseau. Cette activité est considérée comme la signature spécifique d'une connectivité intrinsèque au cortex où les boucles de réverbération sont beaucoup plus nombreuses que les afférences directes (entrées externes) traduisant le message reçu par la périphérie sensorielle. L'irrégularité n'est pas en soi une preuve de stochasticité et de variabilité intrinsèque élevée, comme l'ont démontré des modèles génériques déterministes de réseaux récurrents corticaux. Il n'en reste pas moins vrai que l'efficacité du codage sensoriel dépend à la fois de la variabilité au cours du temps des réponses évoquées et de leur reproductibilité (pour un même signal d'entrée). Comme il est démontré que le neurone est un intégrateur fiable du bombardement synaptique, la vue dominante est de considérer, au niveau du potentiel de membrane comme du processus discret de décharge, un schéma additif entre “bruit” (stochastique) et “signal” (déterministe). Cette approche justifie un code neuronal établi à partir d'un moyennage temporel ou au travers de l'activité collective d'assemblées neuronales. Ce concept a été appliqué pour différentes échelles d'intégration spatio-temporelle, microscopique (enregistrement intracellulaire), mésoscopique (enregistrements optiques avec des colorants sensibles au potentiel) et macroscopiques (imagerie à résonance magnétique fonctionnelle, IRMf). Une vue alternative est de considérer que le “bruit”, en tant que tel, n'existe pas: il est en effet vraisemblable que des stimuli ayant une richesse spectrale temporelle élevée soient requis pour révéler la limite supérieure de la précision temporelle du codage. Il est également possible que les activités évoquées et spontanées interagissent non-linéairement: pour certains signaux d'entrée, l'interaction serait suppressive (shunt inhibiteur) produisant une dégradation dans la propagation du signal. Pour d'autres statistiques externes, une “résonance” au contraire pourrait opérer, menant à l'incrustation du motif évoqué dans l'activité en cours du réseau, et résultant en une propagation sans corruption. Je présenterai des données électrophysiologiques et des prédictions des modèles de mon laboratoire, qui suggèrent que la dynamique du réseau cortical visuel s'adapte aux statistiques des signaux sensoriels: Pour des stimuli de faible dimensionnalité, l'architecture du réseau est quasi cristalline, et composée par des modules répétitifs de colonnes fonctionnelles pavant le champ cortical. Les réponses sensorielles sont denses mais variables et le codage nécessite un moyennage spatial ou temporal. A l'opposé, pour des stimuli de haute dimensionalité, la dynamique corticale devient plus complexe (du type "fumée" selon le titre du livre d'Henri Atlan sur la complexité). Elle pourrait devenir sans bruit (au bord du chaos déterministe) dans deux situations: 1) quand les statistiques d'entrée se rapprochent de celles liées à l'expérience passée, au cours du développement ou de l'apprentissage, ou, 2) quand elles reproduisent des patterns ou trajectoires mémorisées dans la dynamique interne du réseau (les « chants corticaux »). Travail soutenu par le CNRS, l'ANR et la Communauté Européenne (Facets et Brain-i-nets)

Jean Ponce
ENS Computer Science Laboratory (France)
"What is a camera"
Abstract:
I will address in this talk the problem of characterizing a general class of cameras under reasonable, "linear'' assumptions. Concretely, I will use the formalism and terminology of classical projective geometry to model cameras by two-parameter linear families of straight lines---that is, degenerate reguli (rank-3 families) and non-degenerate linear congruences (rank-4 families). This model captures both the general linear cameras of Yu and McMillan and the linear oblique cameras of Pajdla. I will show that it affords a simple classification of all possible geometric camera configurations. From an analytical viewpoint, our model also provides a simple and unified methodology for deriving general formulas for projection and inverse projection, triangulation, and binocular and trinocular geometry. I willl conclude with a brief discussion of ongoing work, where we generalize Pajdla's model of oblique cameras by certain projective maps to all linear cameras, and introduce a natural notion of intrinsic parameters for these cameras, further simplifying the corresponding single- and multi-view formulas. Joint work with Guillaume Batog and Xavier Goaoc at INRIA Nancy Grand Est.

Zhenghyou Zhang
Microsoft Research (USA)
"From Robots to Avatars to Humans"
Abstract:
This talk is a personal retrospect of a few research activities I conducted since joining Microsoft Research (MSR). While at INRIA, under Olivier Faugeras’ leadership and guidance, I worked on 3D vision for robot and Mars Rover navigation. After arriving at MSR, I was interested in developing technologies to build 3D avatars from webcams. The first one was camera calibration. Rather than reimplementing classical algorithms, I developed a calibration technique with a freely moving planar pattern that has since been widely used. To my surprise, I learned that others have referred to it as Zhang’s method. For 3D avatar building, I proposed a model-based modeling approach for robust face reconstruction from only a few feature points. After avatars, I decided to move on to develop technologies for computer-mediated human-human communication and collaboration. I temporarily left computer vision and exclusively focused on speech, and developed a robust speech enhancement technique based on multisensory fusion, combining both air- and bone-conductive microphones.  After four years, I felt ready to do research on both audio and video, and started leading a team on joint audio-visual processing for immersive multimodal remote collaboration. We contributed several key technologies to Microsoft RoundTable, an advanced videoconferencing product shipped in 2007. We are now pushing the frontier of multimedia technology and inventing ways to enable distributed teams to engage effectively and naturally in complex and creative endeavors, transcending space and time.

François Germain
Noesis (France)
"Quelques retombées industrielles des travaux d’Olivier"
Abstract:
Dans une première partie, on rappelle les débuts du traitement d’images à l’INRIA qui coïncide avec l’arrivée d’Olivier et les impulsions qu’il a pu apportées pour appliquer les résultats de la recherche à des problèmes concrets. Une seconde partie donnera quelques éléments sur le parcours de Noesis, une des sociétés dont Olivier est le co‐créateur. Ce sera l’occasion de présenter rapidement quelques applications dans des domaines très variés et de donner quelques éléments sur les tous derniers développements. On conclura par quelques questions ouvertes

Luc Robert
Autodesk (France)
"REALVIZ : Computer Vision for Image Creators"
Abstract:
The ROBOTVIS team led by Olivier Faugeras has developed theory and a number of algorithms targeted at perceiving and reconstructing 3D shapes from images. Since its creation in March 1998, REALVIZ has used this technology as the core of its products and made it available to very diverse industries, including the filmmaking business, professional photographers, architects… In this presentation I will highlight the main developments we have conducted at REALVIZ, from its creation to two years after its acquisition by Autodesk.

Nicholas Ayache
Inria (France)
"Towards Personalized Virtual Patients"
Abstract:
I will show how well designed computational models of organs and pathologies can be personalized from medical images and signals of patients, and how these patient-specific models can then support diagnosis, therapy planning, and even predict treatment outcome.
This will be illustrated by current research applied to cardiovascular and brain diseases conducted at INRIA by the Asclepios project-team with several academic, clinical and industrial partners.

Nassir Navab
Technische Universität München (Germany)
"Modeling, Imaging and Visualization for Computer Assisted Interventions"
Abstract:
In this talk, I will focus on the problem of design and development of advance imaging and visualization solutions for computer assisted interventions. One major scientific challenge is the recovery and modeling of surgical workflow. The second one is the analysis of large amount of heterogeneous data and their intelligent real-time fusion. The third one is the advanced visualization of such data during focused, high-intensity surgical procedures. Throughout this presentation, I use clinical applications and our recent results, obtained in our real-world laboratories within several clinics in Munich, to demonstrate the issues and to provide exemplary paths towards possible solutions. Such examples include real-time Ultrasound/CT registration, Free-Hand SPECT reconstruction, dynamic cone-beam reconstruction, Camera-Augmented Mobile C-arm (CAMC) and HMD based AR for intra-operative visualization and medical training.

Jonathan Touboul
Department of Mathematical Physics of the Rockefeller University (USA)
"Nonlinear Dynamics, Neural Mass Activity and Epilepsy"
Abstract:
Lumped-parameter models of neuronal assemblies, sometimes called neural masses, are important because they can be precisely related to measurements and from there, possibly, to clinical states. The different regimes of a given model, its "œbehaviours"€, can be analytically characterized with the help of dynamical systems analysis, in particular bifurcation theory. These behaviours can in turn be put in correspondence with typical features observed in, e.g., electroencephalographic recordings (EEG). We provide a correspondance between the existence of certain bifurcations in the model and the presence of typical behaviors in the signal. This provides a systematic method to track behaviors in such model, program that we implement in a two populations model, one excitatory, one inhibitory, featuring feedbacks. We completely analyze this model's local and global bifurcations of fixed points and limit cycles of codimension up to three that occur when varying parameters, and relate these mathematical phenomena to typical events in EEG recordings and epilepsy. We analyse also the effect of noise and its relationship with interictal spikes and fast onset activity, a phenomenon of particular relevance in the hippocampus. This provides an extensive panorama of the model behaviours which we discuss with an eye toward their relation with epileptic activity, as reflected in EEG signals. This allows to test hypothesis on the model, such as the link between hyperinnervation (increase synaptic connections) and epilepsy, and to emulate the effect of anticonvulsant medication antagonist of GABA neurotransmitter on the interruption of epileptic seizures.