Journée Machine Learning Optimisation : 13 décembre 2016

Les journées scientifiques SIERRA et LIMA ont décidé de se réunir pour proposer une journée scientifique régionale portant sur le problème de l'Optimisation dans le cadre du Machine Learning. Cette journée aura lieu à Lyon :

Journée scientifique LIMA : Jeudi 8 Octobre 2015

La journée scientifique régionale du projet LIMA (Loisirs et Images) aura lieu :

Inscrivez-vous ici.

Programme

• Cengiz Oztirelli, ETH Zurich
• Gabriel Peyre - Ceremade Paris
• Mathieu Desbrun – Caltech (United States)

• Thierry Chateau, Responsable de l’équipe de Recherche ComSee à l’Institut Pascal

Appel à poster

Programme détaillé

• 9h00 :
  • Accueil


• 9h15 à 10h15 :
  • Cengiz Oztirelli, ETH Zurich, "Point Process Statistics for Computer Graphics"


• 10h15 à 11h15 :
  • Thierry Chateau, Institut Pascal, "Localiser une caméra dans son environnement pour des applications de conduite autonome et de réalité virtuelle"


• 11h15 à 11h45 :
  • Session de présentations "flash"


• 11h45 à 13h15 :
  • Buffet


• 13h15 à 14h00 :
  • Session Poster


• 14h00 à 15h00 :
  • Gabriel Peyre, CNRS et Ceremade, Univ. Paris-Dauphine, "Transport optimal numérique et applications"


• 15h00 à 16h00 :
  • Mathieu Desbrun, Caltech, "Discrete vector field on triangle meshes"


• 16h00 à 16h30 :
  • Message de Patrick Lambert concernant le prochain appel à projets de l'ARC6
  • Message de David Gal Régnier (Pôle de compétitivité Imaginove)


• 16h30 à 16h40 :
  • Clôture de la journée

Présentation des exposés :

• Cengiz Oztirelli, ETH Zurich, "Point Process Statistics for Computer Graphics"

  • Dr. Cengiz Öztireli is a Senior Research Associate in the Computer Graphics Laboratory of Prof. Markus Gross at ETH Zürich. His research interests are in geometry and image processing.

  • Point distributions are ubiquitous in computer graphics. Each time we sample an entity such as a shape or function, we are relying on the properties of the sampling. These properties have so far been studied using spectral tools or simpler measures in computer graphics. We have recently witnessed the utilization of another matured theory specifically designed for point patterns in physics and statistics. The field of point process statistics offers a comprehensive set of theoretical and practical methods, which we can adapt and extend for the problems in computer graphics. In this talk, I will explain the theoretical and practical tools offered by this discipline, and how they can greatly enhance our understanding and algorithms for many problems in computer graphics.

• Thierry Chateau, Institut Pascal, "Localiser une caméra dans son environnement pour des applications de conduite autonome et de réalité virtuelle"

  • Thierry CHATEAU est professeur à l’Institut Pascal où il dirige l’axe Image, Systèmes de Perception et Robotique. Il développe des méthodologies autour de l’interprétation de scènes à partir de vidéos. Il est également vice-président de l’AFRIF (Association Française de Reconnaissance et d’Interprétation des Formes).

  • Cet exposé abordera plusieurs méthodes de type structure from motion pour la reconstruction 3D et la localisation d’un système à base de caméras. L’accent sera mis sur l’aspect temps réel de ces systèmes et sur des applications telles que le guidage autonome de véhicule ou la réalité augmentée. L’exposé sera illustré par plusieurs exemples de réalisations pratiques et de valorisations de ces technologie vers le monde industriel.

• Gabriel Peyre, CNRS et Ceremade, Univ. Paris-Dauphine, "Transport optimal numérique et applications"

  • Gabriel Peyré dirige le projet ERC SIGMA-Vision et travaille sur l'analyse théorique et numérique des problèmes inverses en imagerie. Les outils développés sont issus de l'optimisation convexe non lisse, en particulier les régularisations de faible complexité (parcimonie, variation totale, matrices de faible rang, etc.) et le transport optimal. Il est également le créateur du site web www.numerical-tours.com, qui propose de nombreuses ressources en Matlab/Scilab/Python/Julia pour apprendre et enseigner les mathématiques et l'algorithmique de l'imagerie. Il est directeur de recherche au CNRS affilié au Ceremade, Université Paris-Dauphine.

  • Le transport optimal s'est maintenant imposé comme un outil fondamental pour analyser différents problèmes théoriques à l'interface entre le calcul des variations, les équations aux dérivées partielles et les probabilités. Il a cependant fallu plus de temps pour cette notion devienne également incontournable dans les applications numériques. Ceci est en grande partie du au cout du calcul liés aux problèmes d'optimisation correspondants. On commence cependant à voir émerger de nombreuses applications dans des champs aussi divers que l'astrophysique, la vision par ordinateur, l'infographie, le traitement d'image, les statistiques et l'apprentissage. Dans cet exposé, je ferai un tour d'horizon d'une classe de méthodes rapides pour l'approximation de fonctionnelles de transport optimal. Ces méthodes utilisent une régularisation entropique des problèmes variationnels considérés pour approcher la solution à l'aide d'algorithmes rapides et parallelisable. Ils permettent pour la première fois de calculer des barycentres au sens de la distance de transport d'un grand nombre de densités définis sur des gilles de millions de points. Ceci offre de nouvelles perspectives pour l'application du transport optimal en apprentissage (classification d'histogrammes d'images ou de mots) et en infographie (transfert de couleurs, de textures et de formes). Ces algorithmes permettent aussi de calculer des flots gradients pour la métrique de transport, permettant par exemple d'approximer avec un algorithme rapide le modèle de mouvement de foules proposé par Maury, Roudneff-Chupin et Santambrogio. Il s'agit d'un travail en collaboration avec J.D. Benamou, G. Carlier, M. Cuturi et L. Nenna.

• Mathieu Desbrun, Caltech (United States), actuellement en séjour sabbatique en France à l’INRIA Sophia Antipolis. "Discrete vector field on triangle meshes"

  • Head of the Information Sciences and Mathematics Department of Caltech (United States), actuellement en séjour sabbatique en France à l’INRIA Sophia Antipolis.

  • While scalar fields on surfaces have been staples of geometry processing, the use of tangent vector fields has steadily grown over the last two decades. Tangent vector fields are now a key ingredient to encode directions and sizing intrinsically on surfaces as commonly required in geometry processing tasks such as texture synthesis, non-photorealistic rendering, digital grooming, and meshing. There are, however, a variety of discrete representations of tangent vector fields on triangle meshes, and each approach offers a different trade-off among simplicity, efficiency, and accuracy depending on the targeted application. In this talk, we discuss three main families of discretizations used to design computational tools for tangent vector field processing on triangle meshes: face-based, edge-based, and vertex-based representations. In the process of reviewing the computational tools offered by these representations, we go over a large body of recent developments in vector field processing in the area of discrete differential geometry. In particular, we discuss the importance of defining an appropriate notion of metric connection on triangle meshes. We also discuss the theoretical and practical limitations of each type of discretization, and briefly cover increasingly-common extensions such as n-direction and n-vector fields.

Session de présentations "flash" :

• Sergi Pujades Rocamora (joint work with F. Devernay and B. Goldluecke)
  
  • "Bayesian View Synthesis and Image Based Rendering Principles"
    
    
• Rihab Ben-Ameur (travail conjoint avec D. Coquin et L. Valet)
  
  • "Fusion multimodale pour la reconnaissance d'espèces d'arbres"
    
    
• Even Entem (joint work with Loic Barthe and Marie-Paule Cani)
  
  • "Analysis of organic line drawings"
    
    
• Hugo Lafaye de Micheaux (joint work with Philippe Frey and Christophe Ducottet)
  
  • "Multi-object tracking in long sequences to study bedload transport in mountain rivers"
    
    
• Riadh Ben Salah (travail conjoint avec Olivier Alata, Benoit Tremblais, Lionel Thomas et Laurent David)
  
  • "Reconstruction de volume de particules fondée sur un processus ponctuel marqué et application à la Tomo-PIV"
    
    
• Zhangyun Tan (Joint work with Abdourrahmane Atto, Olivier Alata and Maxime Moreaud)
  
  • "ARFBF model for non stationary random fields and application in HRTEM images"
    
    

Journée scientifique LIMA2 : Mardi 15 avril 2014

La journée scientifique régionale du projet LIMA2 (Loisirs et Images) aura lieu :

Inscrivez-vous ici.

Programme

• 9h15 :
  • Accueil


• 9h30 à 10h30 :
  • Edmond Boyer


• 10h30 à 11h30 :
  • Matthieu Cord


• 11h30 à midi :
  • Session de présentations "flash"


• 12h à 13h30 :
  • Déjeuner


• 13h30 à 14h15 :
  • Session Poster


• 14h15 à 14h 30 :
  • Message de David Gal Régnier


• 14h30 à 15h30 :
  • Mathieu Aubry


• 15h30 à 16h30 :
  • Diane Larlus


• 16h30 à 16h45 :
  • Clôture de la journée et informations de LIMA2

Présentation des exposés :

• Edmond Boyer (INRIA Grenoble, Morpheo)
  
  

  • The INRIA MORPHEO research team is working on the perception of moving shapes using multiple camera systems. Recent progress in computer vision allows to recover dense information on shapes and their motions using visual cues. This brings a rich domain for research investigations on how to model, understand and animate real dynamic shapes. In this talk I will present and discuss recent activities in the team on the recovery of shape models and motion information as well as on how to consistently segment multi-viewpoint images to this purpose.

• Matthieu Cord (LIP6)
  
  
  • Quadruplet-wise constraints for visual metric learning
    

    Metric learning is useful in many Computer Vision applications. The usual way to learn similarities is by considering binary labels (similar vs. not similar) on image pairs. Metrics are learned from training data to minimize dissimilarities between similar pairs, while separating dissimilar ones. Some authors have attempted to go beyond learning metrics from pairwise schemes by considering triplet-wise constraints. Other methods exploit specific rankings between classes, for instance, relative attributes where the classes are ranked with respect to different attributes.
    That is the context on which I will focus in this presentation, but instead of pairwise or triplet-wise techniques, I will investigate relations between quadruplets of images. I will argue that those relationships --- i.e., two images are more similar than two other images --- allow more freedom to specify user-knowledge about the similarity we are trying to learn. Exploiting those quadruplet-wise constraints, I will detail a metric learning framework based on Mahalanobis-like metrics embedded in a convex optimization scheme. Results on image classification, relative attributes and Webpage comparison will be discussed.

• Mathieu Aubry (INRIA Willow Project)
  
  
  • Recovering the 3D model from a single image: mid-level 2D-3D alignment for visual recognition
    

    In this talk I will explore how one can use the increasing quantity of available 3D models to improve image understanding. I will first show how one can reliably align a non-photographic depiction of an architectural site with its 3D model. This is a tremendously difficult task because the appearance and scene structure in the 2D depiction can be very different from the appearance and geometry of the 3D model, for example, due to the specific rendering style, drawing error, age, lighting, or change of seasons, and because the number of possible alignments of the depiction to a large 3D model is huge. I will then show how a similar approach can be used to build an exemplar-based 3D representation of an object category utilizing a large collection of 3D CAD models. I will demonstrate detection and alignment of the "chairs" object category on challenging Pascal VOC 2012 images using a reference library of 1,394 CAD models downloaded from the Internet.

• Diane Larlus (Xerox)
  
  

  • Predicting an Object Location using a Global Image Representation
    

    In this work, we tackle the detection of prominent objects in images as a retrieval task: given a global image descriptor, we find the most similar images in an annotated dataset, and transfer the object bounding boxes. We refer to this approach as data driven detection (DDD), that is an alternative to sliding windows. Previous works have used similar notions but with task-independent similarities and representations, i.e. they were not tailored to the end-goal of localization. In this talk, I will present two contributions: (i) a metric learning algorithm and (ii) a representation of images as object probability maps, that are both optimized for detection. We show experimentally that these two contributions are crucial to DDD, do not require costly additional operations, and in some cases yield comparable or better results than state-of-the-art detectors despite conceptual simplicity and increased speed. As an application of prominent object detection, we improve fine-grained categorization by pre-cropping images with the proposed approach.

Session de présentations "flash" :

• Yoann Baveye
  
  • "LIRIS-ACCEDE: A Video Database for Affective Content Analysis"
    
    
• Jérémy Levallois
  
  • "Parameter-free and Multigrid Convergent Digital Curvature Estimators"
    
    
• Arnaud Lienhard
  
  • "Photo Rating of Facial Pictures based on Image Segmentation"
    
    
• Natalia Neverova
  
  • "Multi-scale spatial and temporal integration for multi-modal gesture recognition"
    
    
• Sergi Pujades Rocamora
  
  • "It Came From Outer Space"
    
    
• Sabin Tiberius Strat
  
  • "Sacs de Mots de trajectoires - application à l'indexation de vidéos à grande échelle"
    
    
• Florent Wachtel
  
  • "Un système d'échantillonnage rapide, adaptatif et avec contrôle spectral, basé sur un pavage de polyhexes"
    "Fast Tile-based Adaptive Sampling With User-Specified Fourier Spectra"
    
    

Journée scientifique LIMA2 : 31 Janvier 2013

La journée scientifique régionale du projet LIMA2 (Loisirs et Images) a eu lieu :

Programme

• 9h :
  • Accueil


• 9h30 à 10h30 :
  • Julie Digne (LIRIS - GeoMod), High resolution point clouds : how to process them?


• 10h30 à 11h30 :
  • Thibault Weise (EPFL, Faceshift), Realtime facial performance capture for animation and gameplay


• 11h30 à 11h45 :
  • Pause


• 11h45 à 12h00 :
  • Information sur les ARCs (Patrick Lambert)


• 12h à 12h15 :
  • Présentation flash des posters


• 12h15 à 14h00 :
  • Buffet et Posters des doctorants


• 14h00 à 15h00 :
  • Adrien Gaidon (LJK- Lear/INRIA), Structured Models for Action Recognition in Real-world Videos


• 15h00 à 16h00 :
  • François Faure (LJK- Imagine/INRIA), Easy Mechanical Simulation using Frame-based Meshless Deformation


• 16h30 à 16h15 :
  • Pause


• 16h15 à 16h45 :
  • Informations Imaginove (David Gal Régnier)


• 16h45 à 17h15 :
  • Conclusion de la journée, discussion sur les prochains événements

Présentation des exposés :

• Julie Digne (LIRIS-GéoMod)
  
  
  • High resolution point clouds : how to process them?

    Over the recent years, the quality of 3d surface acquisition devices has increased dramatically. Using standard reconstruction techniques with these high-resolution point clouds, the final reconstructed surface is usually smoothed, and the small details and textures captured by the scanner are lost. In this talk, we will describe a new framework for processing this type of data. By analyzing the asymptotic behavior of several projection operators, we introduce a very simple projection operator that is consistent with the mean curvature motion. This operator allows for the extension of a well known 2D image analysis theory, scale spaces, to 3D point clouds. In this unified framework a lot of various problems can be solved: point set orientation, surface mesh reconstruction, scan merging and mesh segmentation for high precision data.

• Thibaut Weise (EPFL, Faceshift)
  
  
  • Realtime facial performance capture for animation and gameplay
    

    I n this talk we will discuss the challenges of facial performance capture, and how with the advent of commercially available 3D cameras such as the Microsoft Kinect real-time facial tracking and animation has become possible. I will present the algorithms that are necessary to achieve accurate tracking and animation despite the noisy and low-resolution data of the cameras. The second part of the talk will be about how we took our developed technology to found our spin-off faceshift, show a demo of our professional animation tool, and talk briefly about the SDK that will be release in 2013.

• Adrien Gaidon (LJK- Lear/INRIA)
  
  
  • Structured Models for Action Recognition in Real-world Videos
    

    During this presentation, I will present some of the results of my PhD thesis, namely novel models to recognize broad action categories --- like "opening a door" and "running" --- in real-world video data such as movies and internet videos. In particular, we investigate how an action can be decomposed, what is its discriminative structure, and how to use this information to accurately represent video content. The main challenge we address lies in how to build models of actions that are simultaneously information-rich --- in order tocorrectly differentiate between different action categories --- and robust to the large variations in actors, actions, and videos present in real-world data. Our approach consists in organizing collections of robust local features into structured action representations, for which we propose efficient kernels. We conducted thorough experiments on real- world videos from challenging benchmarks used by the action recognition community. We show that our methods outperform the related state of the art, thus highlighting that using structure information allows for more accurate and robust action recognition in real-world videos.

• François Faure (LJK- Imagine/INRIA)
  
  

  • Easy Mechanical Simulation using Frame-based Meshless Deformation
    

    A new method to simulate deformable objects with heterogeneous material properties and complex geometries using an arbitrary number of control frames is presented. Given a volumetric map of the material properties of an object and a number of control nodes, a distribution of the nodes is computed automatically, as well as the associated shape functions. Reference frames are attached to the nodes, and deformations are applied to the object using linear blend skinning. A continuum mechanics formulation is derived from the displacements and the material properties. We introduce novel material-aware shape functions in place of the tradi- tional radial basis functions used in meshless frameworks. These allow coarse deformation functions to very efficiently resolve non-uniform stiffnesses. Complex models can thus be simulated at high frame rates using a small number of control nodes. We present applications to physical simulation and image registration.

Session de présentations "flash" :

• Yoann Baveye
  
  • "LIRIS-ACCEDE: A Video Database for Affective Content Analysis"
    
    
• Jérémy Levallois
  
  • "Parameter-free and Multigrid Convergent Digital Curvature Estimators"
    
    
• Rizwan Ahmed Khan et al
  
  • "Detection of emotions from videos in controlled environment"
    
    
• Sabin Tiberius Strat et al
  
  • "Automatic Semantic Concept Detection in Generic Videos Using Retina-Enhanced Spatio- Temporal Descriptors"
    
    
• Abdelkader Hamadi et al
  
  • "Utilisation de relations inter-concepts pour l'indexation sémantique des images/vidéos"
    
    
• Dicko Ali Hamadi et al
  
  • "Anatomy Transfer"
    
    
• Nadia Derbas et al
  
  • "Event detection Using motion descriptors"
    
    
• Jean-David Genevaux et al
  
  • "Terrain Generation using Procedural Models based on Hydrology"
    
    
• Hong-Thinh Nguyen et al
  
  • "String of Histograms Edit Distance for Image Classification"