Séminaire de recherche A3SI
(2015-2016)

Lundi 20 juin 2016

Using Visual Data as Its Own Supervision
Alexei Efros, Phillip Isola and Richard Zhang, UC Berkeley
Abstract: Computer vision has recently made great progress through the use of deep learning, trained with large-scale labeled data. However, good labeled data requires expertise and curation, can be expensive to collect, and will always be available in smaller quantities than unlabeled data. Can we discover useful visual representations without the use of explicitly curated labels? To this end, we explore the paradigm of self-supervised learning: using data itself as labels. We will describe several case studies in which we train a neural net to predict one part of a raw sensory signal from another. First, we learn to in-paint missing pixels given their surrounding context. Second, we learn to add missing color to black and white photos. Finally, we learn to predict missing sounds for silent videos. In each case, by learning to predict data from data, our models end up learning visual representations that are broadly useful. We show that these representations encode meaningful information about the semantics and physics of the world, and can be leveraged to aid a variety of downstream recognition tasks.

Jeudi 16 juin 2016

Atelier doctotants

Approche variationnelle pour la segmentation d'images angiographiques
Olivia Miraucourt, LIGM, ESIEE
Abstract: Dans un premier temps, un bref tour d'horizon des méthodes variationnelles utilisées pour la segmentation sera présenté. Nous proposerons un premier modèle qui inclut un a priori de tubularité dans les modèles variationnels de débruitage ROF et TV-L1. Néanmoins, bien que ces modèles permettent de réhausser les vaisseaux dans l'image, ils ne permettent pas de les segmenter. C'est pourquoi nous proposerons un deuxième modèle qui inclut à la fois un a priori de tubularité et un a priori de direction dans le modèle variationnel de segmentation de Chan-Vese. Des résultats sont fournis sur des images synthétiques 2D, ainsi que sur des images rétiniennes de la base DRIVE.

Shallow Packings and Geometry
Bruno Jartoux, LIGM, ESIEE
Abstract: We introduce a combinatorial analogue of Euclidean sphere packing problems: Given a family of sets, one wants to find a largest subfamily whose sets have pairwise high symmetric difference (a packing). When such set families arise from geometry it is often possible to significantly sharpen the upper bound on the cardinality of packings. Haussler's bound on their size (1995) was recently refined (2014-2016). We discuss these results; in particular we prove their tightness by giving a simple construction of packings whose size reaches those bounds.

Exploiting image noise in digital image forensics
Thibault Julliand, LIGM, ESIEE
Abstract: Noise is an intrinsic specificity of all forms of imaging, and can be found in various forms in all domains of digital imagery. We will present an overall review of digital image noise, from its causes and models to the degradations it suffers along the image acquisition pipeline. We will then show how noise can be exploited in image forensics. Image forensics is based on detecting various manipulations on images. Here, we will focus on a specific method called image splicing. Image splicing is a common manipulation which consists in copying part of an image in a second image. Here, we will show how to exploit the variation in noise characteristics in spliced images, caused by the difference in camera and lighting conditions during the image acquisition. Our method automatically gives a probability of alteration for any area of the image, using a local analysis of noise density. We consider both Gaussian and Poisson noise components to modelize the noise in the image.

Mardi 7 juin 2016

Convergent geometric estimators with digital volume and surface integrals
Jacques-Olivier Lachaud, LAMA, Université de Savoie
Abstract: This talk presents several methods to estimate geometric quantities on subsets of the digital space Z^d. We take an interest both on global geometric quantities like volume and area, and on local geometric quantities like normal and curvatures. All presented methods have the common property to be multigrid convergent, i.e. the estimated quantities tend to their Euclidean counterpart on finer and finer digitizations of (smooth enough) Euclidean shapes. Furthermore, all methods rely on digital integrals, which approach either volume integrals or surface integrals along shape boundary. With such tools, we achieve multigrid convergent estimators of volume, moments and area in Z^d, of normals, curvature and curvature tensor in Z^2 and Z^3, and of covariance measure and normals in Z^d even with Hausdorff noise.

Jeudi 19 mai 2016

Atelier doctotants

Shape-Based Analysis on Component-Graphs for Multivalued Image Processing
Eloïse Grossiord, LIGM, ESIEE, KEOSYS
Abstract: Connected morphological operators based on hierarchical image models have been increasingly considered to provide effi cient image segmentation and filtering tools in various application fields, e.g. (bio)medical imaging, astronomy or satellite imaging. Among hierarchical image models, component-trees represent the structure of grey-level images by considering their nested binary level-sets obtained from successive thresholds. Recently, a new notion of component-graph was introduced to extend the component-tree to model any greylevel or multivalued images. The notion of shaping was also recently introduced as a way to improve the antiextensive filtering of grey-level images by considering a two-layer component-tree for grey-level image processing. In this article, we study how component-graphs (that extend the component-tree from a spectral point of view) and shapings (that extends the component-tree from a conceptual point of view) can be associated for the effective processing of multivalued images. The relevance and usefulness of such association are illustrated by applicative examples.

Robust and accurate line-based pose estimation without Manhattan assumptions
Yohann Salaun, LIGM, ENPC, IMAGINE
Abstract: Usual SfM techniques based on feature points have a hard time on scenes with little texture or presenting a single plane, as in indoor environments. Line segments are more robust features in this case. We propose a novel geometrical criterion for two-view pose estimation using lines, that does not assume a Manhattan world. We also define a parameterless (a contrario) RANSAC-like method to discard calibration outliers and provide more robust pose estimations, possibly using points as well when available. Finally, we provide quantitative experimental data that illustrate failure cases of other methods and that show how our approach outperforms them, both in robustness and precision.

A new Human Recognition system based on efficient Optical Disc detection and Retinal Image Ring extraction
Takwa Chihaoui, LIGM, ESIEE, Université de Tunis El Manar
Abstract: Retinal recognition is an attractive topic of scientific research, due to its unicity, universality and robustness. However, existing systems may suffer from some issues. Indeed, due to the retinography acquisition process, retinal images are often affected by imperfections such as poor gray level contrast, noise and background intensity variation. In the other hand, the dense structure of vessels in retina increases the execution time of key point based recognition process and the rate of mismatching individuals. In order to overcome these problems, we propose in this work a new method called "ODR" (Optical Disc interest Ring) which improves the retinal image quality and extracts an interest ring around the detected optical disc. For evaluation, ODR based both identification and verification systems are assessed through SIFT and SURF description and a subset of the VARIA healthy retinal database. Obtained results show the efficiency of our proposed method which allows speeding up the different evaluated systems while ensuring a high accuracy rate (more than 99%) and outperforming existing systems. In addition, more experiments are conducted on medical STARE and DRIVE databases. Promising results are also obtained.

Bijectivity certification of 3D digitized rotations
Kaçper Pluta, LIGM, LAMA, ESIEE
Abstract: Euclidean rotations in Rⁿ are bijective and isometric maps. Nevertheless, they lose these properties when digitized in Zⁿ. For n=2, the subset of bijective digitized rotations has been described explicitly by Nouvel and Rémila and more recently by Roussillon and Coeurjolly. In the case of 3D digitized rotations, the same characterization has remained an open problem. We propose an algorithm for certifying the bijectivity of 3D digitized rational rotations using the arithmetic properties of the Lipschitz quaternions.

Jeudi 7 avril 2016

Atelier doctotants

Parallelization strategy for elementary morphological operators on graphs
Imane Youkana, LIGM, ESIEE
Abstract: This talk focuses on the graph-based mathematical morphology operators presented in [J. Cousty et al, "Morphological filtering on graphs", CVIU 2013]. These operators depend on a size parameter that specifies the number of iterations of elementary dilations/erosions. Thus, the associated running times increase with the size parameter. In this article, we present distance maps that allow us to determine (by thresholding) all considered dilations and erosions. The algorithms based on distance maps allow the operators to be computed with a single linear-time iteration, without any dependence to the size parameter. Then, we investigate a parallelization strategy to compute these distance maps. The idea is to build iteratively the successive level-sets of the distance maps, each level set being traversed in parallel. Under some reasonable assumptions about the graph and sets to be dilated, our parallel algorithm runs in O(n/p + K log2 p) where n,p, and K are the size of the graph, the number of available processors, and the number of distinct level-sets of the distance map, respectively.

Architecture Dynamiquement Auto-Adaptable pour Systèmes de Vision Embarquée Multi-Capteurs
Ali Isavudeen, LIGM, ESIEE, SAGEM
Résumé : Les systèmes de vision embarquée ont enregistré durant ces dernières années une évolution presque révolutionnaire. La technologie des systèmes sur puces a démontré une grande capacité d'intégration, facilitant ainsi le développement de systèmes embarqués à fonctionnalités multiples et variées. Malgré les avancées technologiques, les contraintes des systèmes embarqués (faible encombrement, basse consommation), la complexité des nouvelles applications, l'augmentation de la résolution des capteurs, ainsi que la variété spectrale (couleur, infrarouge, bas niveau de lumière) restent toujours un défi de la communauté scientifique et industrielle. Une des réponses à cette problématique est de mettre en place une exploitation intelligente des ressources matérielles. Cette intelligence peut consister en la capacité de l'architecture matérielle du système à s'auto-adapter dynamiquement au besoin du contexte d'utilisation du système.
Nous nous inspirons du principe des systèmes auto-adaptables conscients pour fonder notre architecture auto-adaptable. Une boucle de supervision reposant sur l'observation, la décision et l'adaptation permet d'intégrer à l'architecture la capacité de s'auto-adapter au cours de l'utilisation du système. Ces différentes unités constituent le Moniteur du système. Les travaux de thèse se penchent d'une part à l'intégration du Moniteur dans l'architecture, d'autre part à l'amélioration de la flexibilité de l'architecture. La mise en place du Moniteur nécessite des réflexions sur l'acheminement des données d'observation mais également sur l'algorithme de décision qui dicte les adaptations à réaliser au niveau de l'architecture de traitement.
Dans un premier temps, nous avons exploré une solution permettant d'adapter l'architecture de traitement à la volée selon le type de capteur. Cette solution ajoute à la technique dite de Reconfiguration Dynamique Partielle, la gestion de l'adaptation par le Moniteur. Dans une seconde piste, nous proposons une adaptation par rapport à la fréquence trames et la résolution du capteur en ajustant la fréquence pixel du traitement. Actuellement nous finalisons des travaux portant sur la mutualisation de ressources de calcul dans le contexte d'une architecture de traitement multi-capteurs hétérogènes. Enfin, un prochain volet abordera l'intégration des fonctionnalités de supervision dans un système d'interconnexion flexible.

Superpixel Convolutional Networks using Bilateral Inceptions
Raghudeep Gadde, LIGM, IMAGINE, ENPC
Abstract: In this work we propose a CNN architecture for image segmentation. We introduce a new "bilateral inception" module that can be inserted in existing CNN architectures and performs bilateral filtering, at multiple feature-scales, between superpixels in an image. The feature spaces for bilateral filtering and other parameters of the module are learned end-to-end using standard backpropagation techniques. The bilateral inception module addresses two issues that arise with general CNN segmentation architectures. First, this module propagates information between (super) pixels while respecting image edges, thus using the structured information of the problem for improved results. Second, the layer recovers a full resolution segmentation result from the lower resolution solution of a CNN. In the experiments, we modify several existing CNN architectures by inserting our inception modules between "fully-connected" (1 × 1 convolution) layers. Empirical results on three different datasets show reliable improvements not only in comparison to the baseline networks, but also in comparison to several dense-pixel prediction techniques such as CRFs, while being competitive in time.

Automated quantification of the epidermal aging process using in-vivo confocal microscopy
Julie Robic, LIGM, ESIEE, Clarins
Abstract: Reflectance confocal microscopy (RCM) is a powerful tool to visualize the skin layers at cellular resolution. The epidermal layer appears as a honeycomb pattern, whose regularity decreases with age. Our aim is to provide a method to automatically quantify the regularity of the honeycomb pattern. The proposed strategy relies on a cell-level supervised classification as "regular" or "irregular" using spatial information given by a prior segmentation. The aggregated scores defined by the classification results show significant correlation with chronological aging and photo-aging. Thus, our method enables practitioners to more objectively assess the quality of the epidermal layers on large cohort of subjects.

Jeudi 24 mars 2016

Surfaces de révolution discrète et généralisation
Eric Andres, XLIM, Université de Poitiers
Résumé : Nous nous proposons de montrer comment l'on peut réaliser des surfaces de révolution discrète où l'on peut contrôler la topologie et choisir des formes plus générales que le simple cercle de révolution. Nous discuterons de la construction de surfaces de révolution discrète à partir de dessin à main levée et de surfaces et courbes plus générales.

Vendredi 18 mars 2016

Multidisciplinary Computational Anatomy for Medical Image Analysis: From Shape to Function and Pathology
Yoshinobu Sato, NAIST, Japan
Abstract: Computational anatomy models statistically represent inter-subject variability of human anatomy including individual organ shapes and their interrelations. These models are regarded as prior probabilities of human anatomy and can be utilized for Bayesian estimation problems related to anatomical identification. This talk presents medical image analysis of the abdominal organs and musculoskeletal structures using the computational anatomy models. In addition, multidisciplinary extensions of computational anatomy will be discussed especially on integration of organ function and pathology.

Jeudi 18 février 2016

On connectivity of discretized explicit 2D curve and 3D surface
Akihiro Sugimoto, NII, Japan
Abstract: Preserving connectivity is an important property commonly required for object discretization. Connectivity of a discretized object differs depending on how to discretize its original object. The morphological discretization is known to be capable of controlling the connectivity of a discretized object, by selecting appropriate structuring elements. The analytical approximation, which approximates the morphological discretization by a finite number of inequalities, on the other hand, is recently introduced to reduce the computational cost required for the morphological discretization. However, whether this approximate discretization has the same connectivity that the morphological discretization has is yet to be investigated. In this paper, we study the connectivity relationship between the morphological discretization and the analytical approximation, focusing on 2D explicit curves. We show that they guarantee the same connectivity for 2D explicit curves/surfaces.

Jeudi 11 février 2016

p-laplaciens sur graphe et applications en traitement d'images et de nuages de points
Abderrahim El Moataz Billah, GREYC, Université de Caen
Résumé : L'opérateur p-laplacien dans ses formulations continues : variationnelles ou non variationnelles, locales ou non locales ou discrètes a été utilisé pour décrire d'importants processus en physique, biologie, économie, etc. Il est utilisé pour modéliser différentes applications allant du traitement d'image, classification de données de grande dimension à la théorie des jeux stochastique. Dans cet exposé, je montrerai que dans le cadre des équations aux différences partielles sur graphes, on peut donner une définition de l'opérateur p-laplacien qui unifie ces différentes formulations. Je donnerai différents exemples de son utilisation pour résoudre des problèmes variés de traitement d'images ou de nuages de points: restauration, "inpainting", colorisation, etc.

Jeudi 19 novembre 2015

Atelier doctotants

Real to rendered-view adaptation for exemplar 2D-3D detection
Francisco Vitor Suzano Massa, LIGM, IMAGINE, ENPC
Abstract: We present an end-to-end approach for 2D-3D exemplar detection with a convolutional neural network (CNN). Our pipeline performs detection by comparing candidate bounding boxes to a potentially large set of 3D models rendered views. We demonstrate that the ability to adapt the features of natural images to better align with those of CAD rendered views is critical to the success of this technique, and that this adaptation can be efficiently learned without any annotated data. We show that our approach naturally fits in a CNN architecture, and brings to 2D-3D exemplar-based detection the benefits from recent advances in deep learning, both in terms of accuracy and speed. We evaluate our method against two key approaches for performing detection by 3D alignment without any image annotation. We show that we clearly out-perform previous state-of-the-art approach for instance detection on several classes from the IKEA benchmark and that we improve by a large margin the results of Aubry et al. for "chair" detection on their subset of Pascal VOC.

Suppression du flicker pour l'acquisition vidéo en haute vitesse
Ali Kanj, LIGM, ESIEE
Résumé : Lors de l'acquisition des vidéos à haute vitesse (au delà de 100 images par seconde), l'utilisation d'un éclairage artificiel peut induire un niveau de luminosité non uniforme et une variation colorimétrique d'une image à l'autre, ce qui est couramment dénommé "flicker" en anglais. Ce problème est devenu plus courant compte tenu de la disponibilité de dispositifs d'acquisition à haute vitesse grand public, par exemple, dans les générations les plus récentes de smartphones ou les caméras "sport/action". L'objet de notre travail est de stabiliser les variations lumineuses dans les séquences d'images et supprimer le flicker en présence de plusieurs sources de lumière qui illuminent la scène et avec des mouvements complexes. Nos travaux résolvent ces problèmes compliqués et servent à stabiliser la luminosité et la colorimétrie d'une manière locale en gérant simultanément le suivi du mouvement par bloc et l'estimation des paramètres de transformation colorimétrique.

Cut Pursuit: a working set algorithm to learn piecewise constant functions
Loïc Landrieu, LIGM, IMAGINE, ENPC
Abstract: Piecewise constant functions, i.e. functions that have a low number of level sets, are ubiquitous in image analysis as well as medical imagery. We propose a working-set algorithm to solve problems penalized by the convex regularization known as total variation for such functions. Our algorithms exploit the piecewise constant structure by recursively splitting the level-sets using graph cuts. We obtain significant speed up on images that can be approximated with few level-sets compared to state-of-the-art algorithms, as well as efficient regularization path computation.

Lundi 12 octobre 2015

Stereo camera based drivable area detection for autonomous driving
Takuya Nanri, Nissan Co., Japan
Abstract: We propose a general-purpose road boundary detection method, which can detect low height curbs, curved curbs, vegetation and ditches. Many conventional methods based on Digital Elevation Map (DEM) can only deal with ordinary curbs because they focus on step-edge detection in DEM. The proposed method adopts a dedicated edge detection process on several multi-directional scanning lines and a temporal filtering process. Therefore, it can detect a variety of road boundaries even with smooth edges and jagged boundaries robustly and precisely. The experimental results show that the proposed method can detect more kinds of road boundaries than a conventional DEM-based approach.

Vendredi 25 septembre 2015

Fast object detection with binary feature representation
Hironobu Fujiyoshi, Chubu University, Japan
Abstract: Object detection involves classification of a huge number of detection windows obtained by raster scanning of an input image. In this talk, we introduce an approximate computation of linear SVM with binary feature representation to object detection. We will show that the proposed method is about 16 times faster than the conventional HOG and linear SVM and improves the classification accuracy by about 6%.