Séminaire de recherche A3SI
(2010-2011)

Jeudi 30 juin 2011 - 13h30 - salle 260

"End-to-end" machine learning of image segmentation (for Connectomics)
Srini Turaga - Sebastian Seung's lab at MIT
Abstract: Connectomics is a new research effort in neuroscience dedicated to mapping the connectivity of real biological neural networks in the brain. It is still early days and much effort is still dedicated to developing good methods for mapping neural networks. And high-throughput, highly-accurate, automatic image segmentation is the most important technology for the success of connectomics. In this talk, I will present a new machine learning method for the supervised learning of image segmentation. Supervised machine learning is a powerful tool for creating accurate image segmentation algorithms that are well adapted to any dataset. Such algorithms have three basic components: 1) a parametrized function for producing segmentations from images, 2) an objective function that quantifies the performance of a segmentation algorithm relative to ground truth, and 3) a means of searching the parameter space of the segmentation algorithms for an optimum of the objective function.
I will also present new work in each of these areas: 1) a segmentation algorithm based on convolutional networks as boundary detectors, 2) the Rand index as a measure of segmentation quality, and 3) the MALIS algorithm, based on ultrametric learning, for training boundary detectors to optimize the Rand index segmentation measure. Taken together, these three pieces constitute the first system for truly "end-to-end" learning of image segmentation, where all parameters in the algorithm are adjusted to directly minimize segmentation error.
References:
http://www.mit.edu/people/sturaga/papers/Jain2010COIN.pdf
http://www.mit.edu/people/sturaga/papers/Turaga2009NIPS.pdf

Vendredi 24 juin 2011 - 13h30 - salle 110

Super-Labels and Hierarchical Label Costs
Andrew Delong - University of Western Ontario
Abstract: The first part of the talk will be about a simple segmentation functional similar to Boykov-Jolly / GrabCut, The standard setup includes a GMM for foreground label, another for the background, and an MRF for regularization. Everyone agrees that the MRF aspect is important, because without it the pixels are assumed i.i.d. and this is simply not a good model for natural images. However, if the i.i.d. assumption is inappropriate for generating an image, then it is equally inappropriate for generating complex objects *within* the image (person, car), and yet this is what the standard setup entails. I'll talk about how to do better using a "2-level MRF" defined over "super-labels." The second part of the talk will present some unpublished work on energies with a hierarchy of "label costs" and how to optimize them effectively. Such energies facilitate a kind of "hierarchical MDL" criterion and should be useful in detecting multiple objects, motions, homographies, and much more. I'll briefly demonstrate a way to detect repetitive patterns using this framework.

Jeudi 16 juin 2011

Geodesic methods for Biomedical Image Segmentation
Laurent D. Cohen - CEREMADE, UMR CNRS 7534, Université Paris Dauphine
Abstract: Tubular and tree structures appear very commonly in biomedical images like vessels, microtubules or neuron cells. Minimal paths have been used for long as an interactive tool to segment these structures as cost minimizing curves. The user usually provides start and end points on the image and gets the minimal path as output. These minimal paths correspond to minimal geodesics according to some adapted metric. They are a way to find a (set of) curve(s) globally minimizing the geodesic active contours energy. Finding a geodesic distance can be solved by the Eikonal equation using the fast and efficient Fast Marching method. In the past years we have introduced different extensions of these minimal paths that improve either the interactive aspects or the results. For example, the metric can take into account both scale and orientation of the path. This leads to solving an anisotropic minimal path in a 2D or 3D+radius space. On a different level, the user interaction can be minimized by adding iteratively what we called the keypoints, for example to obtain a closed curve from a single initial point. The result is then a set of minimal paths between pairs of keypoints. This can also be applied to branching structures in both 2D and 3D images. We also proposed different criteria to obtain automatically a set of end points of a tree structure by giving only one starting point. More recently, we introduced a new general idea that we called Geodesic Voting or Geodesic Density. The approach consists in computing geodesics between a given source point and a set of points scattered in the image. The geodesic density is defined at each pixel of the image as the number of geodesics that pass over this pixel. The target structure corresponds to image points with a high geodesic density. We will illustrate different possible applications of this approach. The work we will present involved as well F. Benmansour, Y. Rouchdy and J. Mille at CEREMADE.

Vendredi 10 juin 2011 - 13h30 - salle 260

Geometric Data: Analysis, Optimization and Visualization
Nabil Mustafa
Abstract: We are given some geometric data in the form of a set P of n points in R^d. Then the aim of combinatorial data approximation is to give a succinct summary of the shape of P (e.g., spread and distribution of the points in P). This becomes especially important as the data becomes unmanageably large (e.g., coming from a live-stream), and so only a small 'sketch' can be stored. In this talk I will present various algorithms and techniques for approximating a set of points with a smaller set. I will first present and explain the notion of a centerpoint, which is the generalization of the concept of a median to higher dimensions. I will then present fast algorithms to compute the spread of the pointset based on this notion. A different method to approximate P is via the notion of an epsilon-net, which is a subset of P that approximates P with respect to geometric objects. A well-known theorem states that the existence of small-sized epsilon-nets for P imply constant-factor approximation algorithms for geometric hitting-set problems for P. However, the main limitation of this technique is the inability to give polynomial-time approximation schemes for these problems. I will present a different approach, based on the idea of local-search, that is able to bypass this barrier, and give PTAS for several geometric hitting-set problems.

Jeudi 9 juin 2011 - 13h30 - salle 260

Introduction aux Digital Level Layers et application au calcul des dérivées discrètes
Yan Gérard - ISIT, Université d'Auvergne
Résumé : L'expression Digital Level Layers (DLL) fait référence aux "Level Sets", c'est-à-dire aux ensemble définis par une équation f(x)=0 où la variable x est le plus souvent considérée dans un espace vectoriel réel R^d. Nous lui avons adjoint l'adjectif "Digital" car nous considérons des x dans Z^d plutôt que dans un espace continu et nous avons remplacé "Set" par "Layers" car l'équation f(x)=0 est relaxée en une double-inégalité h<f(x)<h'. Les droites, plans, hyperplans et certaines sphères discrètes classiques sont des DLL avec une fonction f polynomiale de degré 1 ou 2. Nous expliquerons en quoi les DLL sont une alternative intéressante aux autres primitives discrètes définies de façon topologique ou morphologique, en développant quelques idées sur l'algorithmique utilisée pour les reconnaître (ces algorithmes sont aussi bien en relation avec la programmation linéaires, les SVM qu'avec des algorithmes de géométrie algorithmique). Nous verrons aussi comment ce type d'objets fournit une méthode de calcul des dérivées d'une fonction de Z dans Z vérifiant le critère de convergence multi-grilles avec une erreur bornée uniformément (en d'autres termes, une méthode concurrente des convolutions avec noyaux binomiaux).

Jeudi 26 mai 2011

Modèles Booléens multi-échelles pour simuler des microstructures complexes
Dominique Jeulin - Centre de Morphologie Mathématique (Fontainebleau), Mathématiques et Systèmes, Centre des Matériaux P.M. Fourt, Mines ParisTech
Résumé : Les microstructures complexes rencontrées dans les matériaux impliquent souvent des textures hétérogènes multi-échelles, qui peuvent être modélisées par des ensembles aléatoires développés en Morphologie Mathématique. Notre approche s'appuie sur des images 2D ou 3D. Une caractérisation morphologique complète est utilisée pour l'identification d'un modèle de structure aléatoire. Dans cette présentation, nous considérons des modèles d'ensembles aléatoires booléens construits à partir de processus ponctuels multi-échelles de Cox, pour reproduire des microstructures présentant des hétérogénéités à différentes échelles. L'effet de ces échelles multiples sur le comportement physique de milieux hétérogènes est illustré par des applications à des milieux réels.

Jeudi 5 mai 2011

Segmentation and quantification of fluorescence microscopy images
Petr Matula - Masaryk University, Brno
Abstract: In my talk I want to present selected topics related to segmentation and quantification of subcellular structures in fluorescence microscopy images from two projects I have been working on: (1) development of an automated system for the single-cell-based quantification of the level of viral infection using a high-throughput siRNA screening and (2) quantification of the number of fluorescent spots in 3-D confocal fluorescent images of live cells over time. In particular, I want to focus on those parts where we successfully applied mathematical morphology operators and to show images where topology-based operators could hopefully help.

Jeudi 17 mars 2011

Atelier doctotants

Illumination et rendu temps réel de milieux participants hétérogènes par propagation des occlusions
Anthony Giroud - A3SI/LIGM/UPEMLV
Résumé : Nous présentons une nouvelle approche pour l'illumination et le rendu des effets dus à la simple diffusion de la lumière au sein de milieux participants hétérogènes en temps réel. La densité du milieu est dans un premier temps modélisée dans une base de fonctions radiales, puis est ensuite échantillonnée sur une grille volumétrique. Nous intégrons ensuite la fonction d'extinction du milieu entre chaque lumière et chaque case de la grille grâce à un algorithme itératif de propagation radiale sur GPU. Nous visualisons enfin le milieu participant ainsi que les surfaces ombrées via un ray-marching constant depuis l'observateur jusqu'à la plus proche surface, et approximons ainsi une solution à l'équation de diffusion.

Cognitive Radio Spectrum Evolution Prediction using Artificial Neural Networks based Multivariate Time Series Modelling
Muhammad Imran Taj - A3SI/LIGM/ESIEE
Abstract: Cognitive Radio (CR) is an efficient answer to spectrum scarcity as it can sense the spectrum steadily based on previous information about the spectrum evolution in time, thus predicting the future occupancy status. Framed within this statement, this paper proposes a new methodology for spectrum prediction by modelling licensed signal Radio Frequency (RF) features as a multivariate chaotic time series, which is then given as input to Artificial Neural Network (ANN), that predicts the evolution of RF time series to decide if the unlicensed user can exploit the spectrum band. We exploit the inherent cyclostationarity in primary signals for Non-linear Autoregressive Exogenous Time Series Modelling of RF features, which is an extremely challenging task due to interdependence of different RF features. Experimental results show a similar trend between predicted and observed values. We target this work at cognition incorporation in our designed Software Defined Radio (SDR) waveform.

Jeudi 20 janvier 2011

Atelier doctotants

Some morphological operators on simplicial complex spaces
Fabio Augusto Salve Dias - A3SI/LIGM
Abstract: In this work, we propose a framework that allows to build morphological operators for processing and filtering objects defined on (abstract) simplicial complex spaces. We illustrate with applications to mesh and image processing, for which, on the provided examples, the proposed approach outperforms the classical one.

Multichannel image quantization under spatial smoothness constraints
Anna Jezierska - Signal et communications/A3SI/LIGM
Abstract: Quantization, defined as the act of attributing a finite number of levels to an image, is an essential task in image acquisition and coding. It is also intricately linked to essential image analysis tasks, such as denoising and segmentation. We investigate vector quantization combined with regularity constraints, a little-studied area that is of interest, in particular, when quantizing in the presence of noise or other acquisition artifacts. We present an optimization approach to the problem involving a novel two-step, iterative, flexible, joint quantizing-regularization method featuring both convex and combinatorial optimization techniques. We show that for both color and grey-scale images, our approach may yield better quality images in terms of SNR, with lower entropy, than conventional optimal quantization methods.

Efficient parallel, streaming implementation of morphological filters
Jan Bartovsky - A3SI/LIGM
Abstract: The high computation intensity of some advanced morphological operators such as Alternating Sequential Filters (ASF) or granulometries can become a limiting factor of their usability in time-critical applications. We will present a new pipeline HW architecture for fast 2-D erosions/dilations. It uses a recently proposed algorithm allowing to process 2-D data in a stream, minimizing the use of memory and drastically reducing the computing latency. The architecture allows to combine dilations and erosions in an efficient pipeline to compute compound morphological operators with no intermediate image storage and minimal latency. We show how the dedicated implementation on an FPGA achieves a previously unequaled performance, opening an opportunity to use these operators in high-end, real-time applications.

Jeudi 13 janvier 2011

AntiAliasing Morphologique et Topologique
Venceslas Biri - LIGM/A3SI/UPEMLV
Résumé : L'antialiasing morphologique est une approche de l'anticrénelage en post-traitement qui ne requiert pas le rendu d'échantillons supplémentaires. Malheureusement, cet algorithme est un filtre non linéaire, très peu adapté aux architectures massivement parallèles. Nous avons donc repensé la méthode initiale en utilisant plusieurs passes, notamment une nouvelle méthode de calcul des longueurs de lignes, pour un portage sur GPU. De plus, nous introduisons dans la méthode la notion de reconstruction topologique destinée à pallier certaines faiblesses de l'algorithme. Les temps de calculs courts permettent son utilisation en temps réel, même sur des images en haute résolution.

Mardi 7 décembre 2010

Soutenance de thèse : Applications de la topologie discrète pour la captation de mouvement en temps réel et sans marqueurs
Benjamin Raynal - LIGM/A3SI, ESIEE-UPEMLV
Résumé

Jeudi 2 décembre 2010

Soutenance de thèse : Utilisation de la topologie pour l'analyse de formes discrètes
John Chaussard - LIGM/A3SI, ESIEE-UPEMLV
Résumé

Jeudi 25 novembre 2010 - 13h30 - salle 160

Efficient robust digital hyperplane fitting with bounded error
Dror Aiger - LIGM/A3SI/ESIEE
Abstract: We consider the following fitting problem: given an arbitrary set of N points in a bounded grid in dimension d, find a digital hyperplane that contains the largest possible number of points. We first observe that the problem is 3SUM-hard in the plane, so that it probably cannot be solved exactly with computational complexity better than O(N²), and it is conjectured that optimal computational complexity in dimension d is in fact O(N^d). We therefore propose two approximation methods featuring linear time complexity. As the latter one is easily implemented, we present experimental results that show the runtime of the method in practice.
Joint work with Yukiko Kenmochi, Hugues Talbot and Lilian Buzer.

Mardi 16 novembre 2010

Optimization for Pixel Labeling Problems With Structured Layout
Olga Veksler - University of Western Ontario
Abstract: Pixel labeling problems are pervasive in computer vision research. In this talk, we discuss optimization approaches for labeling problems which have some structure imposed on the layout of the labels. In other words, the relationships between labels is not arbitrary but has a well defined spatial structure. We will describe two approaches for structured layout scenes. The first approach is for a more restrictive type of scenes, for which we develop new graph-cut moves which we call order-preserving. The advantage of order preserving moves is that they act on all labels simultaneously, unlike the popular expansion algorithm, and, therefore, escape local minima more easily. The second approach is for a more general type of structured layout scenes and it is based on dynamic programming. In the second case, the exact minimum can be found efficiently. This is very rare for a 2D labeling problem to have an efficient and global optimizer. For both approaches, our applications include geometric class labeling and segmentation with a shape prior.

Mardi 16 novembre 2010

Energy Minimization with Label costs and Applications in Multi-Model Fitting
Yuri Boykov - University of Western Ontario
Abstract: The a-expansion algorithm has had a significant impact in computer vision due to its generality, effectiveness, and speed. Until recently, it could only minimize energies that involve unary, pairwise, and specialized higher-order terms. We propose an extension of a-expansion that can simultaneously optimize "label costs" with certain optimality guarantees. An energy with label costs can penalize a solution based on the set of labels that appear in it. The simplest special case is to penalize the number of labels in the solution, but the proposed energy is significantly more general than this. Usefulness of label costs is demonstrated by a number of specific applications in vision that appeared in the last year. Our work (see CVPR 2010, IJCV submission) studies label costs from a general perspective, including discussion of multiple algorithms, optimality bounds, extensions, and fast special cases (e.g. UFL heuristics). In this talk we focus on natural generic applications of label costs is multi-model fitting and demonstrate several examples: homography detection, motion segmentation, unsupervised image segmentation, compression, and FMM. We also discuss a method for effective exploration of the continuum of labels - an important practical obstacle for a-expansion in model fitting. We discuss why our optimization-based approach to multi-model fitting is significantly more robust than standard extensions of RANSAC (e.g. sequential RANSAC) currently dominant in vision.

Mardi 19 octobre 2010

Soutenance de thèse : Filtrage d'objets fins : applications à l'analyse d'images vasculaires
Olena Tankyevych - LIGM/A3SI, ESIEE-UPEMLV
Résumé
14h00, amphi 160, ESIEE

Jeudi 14 octobre 2010

Soutenance de HDR : Agrégation PAC-Bayésienne et bandits à plusieurs bras
Jean-Yves Audibert - LIGM/A3SI, ENPC/IMAGINE
Résumé
14h00, M209: aile Maupertuis du bâtiment ENSG/ENPC

Mercredi 22 septembre 2010

Soutenance de thèse : Rotations dans les espaces discrets 2D et 3D
Yohan Thibault - LIGM/A3SI, ESIEE-UPEMLV
Résumé