As we navigate the world, for example when driving a car from our home to the work place, we continuously perceive the 3D structure of our surroundings and intuitively recognise the objects we see. Such capabilities help us in our everyday lives and enable free and accurate movement even in completely unfamiliar places. We largely take these abilities for granted, but for robots, the task of understanding large outdoor scenes remains extremely challenging. In this thesis, I develop novel algorithms for (near) real-time dense 3D reconstruction and semantic segmentation of large-scale outdoor scenes from passive cameras. Motivated by "smart glasses" for partially sighted users, I show how such modeling can be integrated into an interactive augmented reality system which puts the user in the loop and allows her to physically interact with the world to learn personalized semantically segmented dense 3D models. In the next part, I show how sparse but very accurate 3D measurements can be incorporated directly into the dense depth estimation process and propose a probabilistic model for incremental dense scene reconstruction. To relax the assumption of a stereo camera, I address dense 3D reconstruction in its monocular form and show how the local model can be improved by joint optimization over depth and pose. The world around us is not stationary. However, reconstructing dynamically moving and potentially non-rigidly deforming texture-less objects typically require "contour correspondences" for shape-from-silhouettes. Hence, I propose a video segmentation model which encodes a single object instance as a closed curve, maintains correspondences across time and provide very accurate segmentation close to object boundaries. Finally, instead of evaluating the performance in an isolated setup (IoU scores) which does not measure the impact on decision-making, I show how semantic 3D reconstruction can be incorporated into standard Deep Q-learning to improve decision-making of agents navigating complex 3D environments.

Bundle adjustment is the process of simultaneously optimizing camera poses and 3D structure given image point tracks. In structure-from-motion, it is typically used as the final refinement step due to the nonlinearity of the problem, meaning that it requires sufficiently good initialization. Contrary to this belief, recent literature showed that useful solutions can be obtained even from arbitrary initialization for fixed-rank matrix factorization problems, including bundle adjustment with affine cameras. This property of wide convergence basin of high quality optima is desirable for any nonlinear optimization algorithm since obtaining good initial values can often be non-trivial. The aim of this thesis is to find the key factor behind the success of these recent matrix factorization algorithms and explore the potential applicability of the findings to bundle adjustment, which is closely related to matrix factorization. The thesis begins by unifying a handful of matrix factorization algorithms and comparing similarities and differences between them. The theoretical analysis shows that the set of successful algorithms actually stems from the same root of the optimization method called variable projection (VarPro). The investigation then extends to address why VarPro outperforms the joint optimization technique, which is widely used in computer vision. This algorithmic comparison of these methods yields a larger unification, leading to a conclusion that VarPro benefits from an unequal trust region assumption between two matrix factors. The thesis then explores ways to incorporate VarPro to bundle adjustment problems using projective and perspective cameras. Unfortunately, the added nonlinearity causes a substantial decrease in the convergence basin of VarPro, and therefore a bootstrapping strategy is proposed to bypass this issue. Experimental results show that it is possible to yield feasible metric reconstructions and pose estimations from arbitrary initialization given relatively clean point tracks, taking one step towards initialization-free structure-from-motion.

Fluorescence microscopy is an indispensable tool in the areas of cell biology, histology and material science as it enables non-invasive observation of specimen in their natural environment. The main advantage of fluorescence microscopy is that, it is non-invasive and capable of imaging with very high contrast and visibility. It is dynamic, sensitive and allows high selectivity. The specificity and sensitivity of antibody-conjugated probes and genetically-engineered fluorescent protein constructs allows the user to label multiple targets and the precise location of intracellular components. However, its spatial reso- lution is limited to one-quarter of the excitation wavelength (Abbe’s diffraction limit). The advent of new and sophisticated optics and availability of fluorophores has made fluorescence imaging a flourishing field. Several advanced techniques like TIRF, 4PI, STED, SIM, SPIM, PALM, fPALM, GSDIM and STORM, have enabled high resolution imaging by breaking the diffraction barrier and are a boon to medical and biological research. Invention of confocal and multi-photon microscopes have enabled observation of the specimen embedded at depth. All these advances in fluorescence microscopy have made it a much sought-after technique. The first chapter provides an overview of the fundamental concepts in fluorescence imag- ing. A brief history of emergence of the field is provided in this chapter along with the evolution of different super-resolution microscopes. An introduction to the concept of fluorophores, their broad classification and their characteristics is discussed in this chap- ter. A brief explanation of different fluorescence imaging techniques and some trending techniques are introduced. This chapter provides a thorough foundation for the research work presented in the thesis. Second chapter deals with different microscopy techniques that have changed the face of biophotonics and nanoscale imaging. The resolution of optical imaging systems are dictated by the inherent property of the system, known as impulse response or more popularly “point spread function”. A basic fluorescence imaging system is presented in this chapter and introduces the concept of point spread function and resolution. The introduction of confocal microscope and multi-photon microscope brought about improved optical sectioning. 4PI microscopy technique was invented to improve the axial resolution of the optical imaging system. Using this microscopy modality, an axial resolution of upto ≈ 100nm was made possible. The basic concepts of these techniques is provided in this chapter. The chapter concludes with a discussion on some of the optical engineering techniques that aid in improved lateral and axial resolution improvements and then we proceed to take on these engineering techniques in detail in the next chapter. Introduction of spatial masks at the back aperture of the objective lens results in gen- eration of a Bessel-like beam, which enhances our ability to see deeper inside a spec- imen with reduced aberrations and improved lateral resolution. Bessel beams have non-diffracting and self-reconstructing properties which reduces the scattering while ob- serving cells embedded deep in a thick tissue. By coupling this with the 4PI super- resolution microscopy technique, multiple excitation spots can be generated along the optical axis of the two opposing high-NA objective lenses. This technique is known as multiple excitation spot optical (MESO) microscopy technique. It provides a lateral resolution improvement upto 150nm. A detailed description of the technique and a thorough analysis of the polarization properties is discussed in chapter 3. Chapters 4 and 5 bring the focus of the thesis to the main topic of research - multi- dimensional image reconstruction for fluorescence microscopy by employing the statis- tical techniques. We begin with an introduction to filtering techniques in Chapter 4 and concentrate on an edge-preserving denoising filter: Bilateral Filter for fluorescence microscopy images. Bilateral filter is a non-linear combination of two Gaussian filters, one based on proximity of two pixels and the other based on the intensity similarity of the two. These two sub-filters result in the edge-preserving capability of the filter. This technique is very popular in the field of image processing and we demonstrate the application of the technique for fluorescence microscopy images. The chapter presents a through description of the technique along with comparisons with Poisson noise mod- eling. Chapters 4 and 5 provide a detailed introduction to statistical iterative recon- struction algorithms like expectation maximization-maximum likelihood (EM-ML) and maximum a-posteriori (MAP) techniques. The main objective of an image reconstruc- tion algorithm is to recover an object from its noisy degraded images. Deconvolution methods are generally used to denoise and recover the true object. The choice of an appropriate prior function is the crux of the MAP algorithm. The remaining of chapter 5 provides an introduction to different potential functions. We show some results of the MAP algorithm in comparison with that of ML algorithm. In chapter 6, we continue the discussion on MAP reconstruction where two new potential functions are introduced and demonstrated. The first one is based on the application of Taylor series expansion on the image. The image field is considered to be analytic and hence Taylor series produces an accurate estimation of the field being reconstructed. The second half of the chapter introduces an interpolation function to approximate the value of a pixel in its neighborhood. Cubic B-splines are widely used as a basis function during interpolation and they are popular technique in computer vision and medical imaging techniques. These novel algorithms are tested on di_erent microscopy data like, confocal and 4PI. The results are shown at the _nal part of the chapter. Tagging cell organelles with uorescent probes enable their visualization and analysis non-invasively. In recent times, it is common to tag more than one organelle of interest and simultaneously observe their structures and functions. Multicolor uorescence imaging has become a key technique to study speci_c processes like pH sensing and cell metabolism with a nanoscale precision. However, this process is hindered by various problems like optical artifacts, noise, autouorescence, photobleaching and leakage of uorescence from one channel to the other. Chapter 7 deals with an image reconstruction technique to obtain noise-free and distortion-less data from multiple channels when imaging a multicolor sample. This technique is easily adaptable with the existing imaging systems and has potential application in biological imaging and biophysics where multiple probes are used to tag the features of interest. The fact that the lateral resolution of an optical system is better than the axial resolution is well known. Conventional microscopes focus on cells that are very close to the cover-slip or a few microns into the specimen. However, for cells that are embedded deep in a thick sample (ex: tissues), it is di_cult to visualize them using a conventional microscope. A number of factors like, scattering, optical aberrations, mismatch of refractive index between the objective lens and the mounting medium and noise, cause distortion of the images of samples at large depths. The system PSF gets distorted due to di_raction and its shape changes rapidly at large depths. The aim of chapter 8 is to introduce a technique to reduce distortion of images acquired at depth by employing image reconstruction techniques. The key to this methodology is the modeling of PSF at large depths. Maximum likelihood technique is then employed to reduce the streaking e_ects of the PSF and removes noise from raw images. This technique enables the visualization of cells embedded at a depth of 150_m. Several biological processes within the cell occur at a rate faster than the rate of acquisition and hence vital information is missed during imaging. The recorded images of these dynamic events are corrupted by motion blur, noise and other optical aberrations. Chapter 9 deals with two techniques that address temporal resolution improvement of the uorescence imaging system. The _rst technique focuses on accelerating the data acquisition process. This includes employing the concept of time-multiplexing to acquire sequential images from a dynamic sample using two cameras and generating multiple sheets of light using a di_raction grating, resulting in multi-plane illumination. The second technique involves the use of parallel processing units to enable real-time image reconstruction of the acquired data. A multi-node GPU and CUDA architecture effciently reduce the computation time of the reconstruction algorithms. Faster implementation of iterative image reconstruction techniques can aid in low-light imaging and dynamic monitoring of rapidly moving samples in real time. Employing rapid acquisition and rapid image reconstruction aids in real-time visualization of cells and have immense potential in the _eld of microbiology and bio-mechanics. Finally, we conclude the thesis with a brief section on the contribution of the thesis and the future scope the work presented. Thank you for using www.freepdfconvert.com service! Only two pages are converted. Please Sign Up to convert all pages. https://www.freepdfconvert.com/membership

La reconstruction 3D d’objets à partir de plusieurs images est un objectif important de la vision par ordinateur. Elle a été largement étudiée pour les objets rigides et non rigides (ou déformables). Le Structure-from-Motion (SfM) est un algorithme qui effectue la reconstruction 3D d’objets rigides en utilisant le mouvement visuel entre plusieurs images obtenues à l’aide d’une caméra en mouvement. Le SfM est une solution très précise et stable. La reconstruction 3D déformable a été largement étudiée pour les images monoculaires (obtenues à partir d’une seule caméra) mais reste un problème ouvert. Les méthodes actuelles exploitent des indices visuels tels que le mouvement visuel inter-image et l’ombrage afin de construire un algorithme de reconstruction. Cette thèse se concentre sur l’utilisation du mouvement visuel inter-image pour résoudre ce problème. Deux types de scénarios existent dans la littérature : 1) le Non-Rigid Structure-from-Motion (NRSfM) et 2) le Shape-from-Template (SfT). L’objectif du NRSfM est de reconstruire plusieurs formes d’un objet déformable tel qu’il apparaît dans plusieurs images, alors que le SfT (également appelé reconstruction à partir d’un modèle de référence) utilise une seule image d’un objet déformé et son modèle 3D de référence (une forme 3D texturée de l’objet dans une configuration) pour estimer la forme déformée de l’objet. (...) / Reconstructing the 3D shape of objects from multiple images is an important goal in computer vision and has been extensively studied for both rigid and non-rigid (or deformable) objects. Structure-from-Motion (SfM) is an algorithm that performs the 3D reconstruction of rigid objects using the inter-image visual motion from multiple images obtained from a moving camera. SfM is a very accurate and stable solution. Deformable 3D reconstruction, however, has been widely studied for monocular images (obtained from a single camera) and still remains an open research problem. The current methods exploit visual cues such as the inter-image visual motion and shading in order to formalise a reconstruction algorithm. This thesis focuses on the use of the inter-image visual motion for solving this problem. Two types of scenarios exist in the literature: 1) Non-Rigid Structure-from-Motion (NRSfM) and 2) Shape-from-Template (SfT). The goal of NRSfM is to reconstruct multiple shapes of a deformable object as viewed in multiple images while SfT (also referred to as template-based reconstruction) uses a single image of a deformed object and its 3D template (a textured 3D shape of the object in one configuration) to recover the deformed shape of the object. We propose an NRSfM method to reconstruct the deformable surfaces undergoing isometric deformations (the objects do not stretch or shrink under an isometric deformation) using Riemannian geometry. This allows NRSfM to be expressed in terms of Partial Differential Equations (PDE) and to be solved algebraically. We show that the problem has linear complexity and the reconstruction algorithm has a very low computational cost compared to existing NRSfM methods. This work motivated us to use differential geometry and Cartan’s theory of connections to model NRSfM, which led to the possibility of extending the solution to deformations other than isometry. In fact, this led to a unified theoretical framework for modelling and solving both NRSfM and SfT for various types of deformations. In addition, it also makes it possible to have a solution to SfT which does not require an explicit modelling of deformation. An important point is that most of the NRSfM and SfT methods reconstruct the thin-shell surface of the object. The reconstruction of the entire volume (the thin-shell surface and the interior) has not been explored yet. We propose the first SfT method that reconstructs the entire volume of a deformable object.

La création d'informations 3D denses à partir d'images endoscopiques intraopératoires, ainsi que le calcul de la position relative de la caméra endoscopique, sont des éléments fondamentaux pour un guidage de qualité durant la chirurgie guidée par l'image. Par exemple, cela permet de superposer modèle pré-opératoire via la réalité augmentée. Cette thèse présente une approche pour l'estimation de ces deux information basées sur une approche de localisation et cartographie simultanées (SLAM). Nous découplons la reconstruction dense de l'estimation de la trajectoire de la caméra, aboutissant à un système qui combine la précision du SLAM, et une reconstruction plus complète. Les solutions proposées dans cette thèse ont été validées sur de séquences porcines provenant de différents ensembles de données. Ces solutions n'ont pas besoin de matériel de suivi externe ni d'intervention. Les seules entrées nécessaires sont les trames vidéo d'un endoscope monoculaire. / Recovering dense 3D information from intra-operative endoscopic images together with the relative endoscope camera pose are fundamental blocks for accurate guidance and navigation in image-guided surgery. They have several important applications, e.g., augmented reality overlay of pre-operative models. This thesis provides a systematic approach for estimating these two pieces of information based on a pure vision Simultaneous Localization And Mapping (SLAM). We decouple the dense reconstruction from the camera trajectory estimation, resulting in a system that combines the accuracy and robustness of feature-based SLAM with the more complete reconstruction of direct SLAM methods. The proposed solutions in this thesis have been validated on real porcine sequences from different datasets and proved to be fast and do not need any external tracking hardware nor significant intervention from medical staff. The sole input is video frames of a standard monocular endoscope.

This work describes a method based on three stages for reconstructing a model from a given set of scanned meshes obtained from 3D scanners. Meshes scanned from different scanner s view points have their representation in local coordinate systems. Therefore, for final model reconstruction, an alignment of the meshes is required. The most popular algorithm for cloud data registration is the ICP algorithm. However, ICP requires an initial estimate of mesh alignment, which is, many times, done manually. To automate this process, this work uses a surface representation called spin-images to identify overlap areas between the meshes and to estimate their alignment. After this initial registration, the alignment is refined by the ICP algorithm, and finally the model is reconstructed using a method called VRIP. / Fundação de Amparo a Pesquisa do Estado de Alagoas / Este trabalho descreve um método baseado em três etapas para reconstrução de modelos a partir de malhas capturadas de scanners 3D. Malhas obtidas a partir de diferentes pontos de visão de um scanner têm sua representação em sistemas de coordenadas local. Portanto, para a reconstrução final do modelo, é necessário realizar um alinhamento dessas malhas, ou registro. O algoritmo mais famoso para realizar registro de nuvens de pontos é o algoritmo ICP. Porém, um dos requisitos desse algoritmo é uma estimativa inicial do alinhamento das malhas, que muitas vezes é feita manualmente. Para automatizar esse processo, este trabalho utiliza descritores spin-image para identificar regiões de sobreposição entre as malhas e estimar seus alinhamentos. Após este registro inicial, o alinhamento é refinado através do algoritmo ICP, e finalmente o modelo é reconstruído usando uma técnica chamada VRIP.

L’objectif de ce travail est d’obtenir un modèle 3D d’un objet à partir d’une série d’images prisesavec un Microscope Electronique à Balayage (MEB). Pour cela, nous utilisons la technique dereconstruction 3D qui est une application bien connue du domaine de vision par ordinateur.Cependant, en raison des spécificités de la formation d’images dans le MEB et dans la microscopieen général, les techniques existantes ne peuvent pas être appliquées aux images MEB. Lesprincipales raisons à cela sont la projection parallèle et les problèmes d’étalonnage de MEB entant que caméra. Ainsi, dans ce travail, nous avons développé un nouvel algorithme permettant deréaliser une reconstruction 3D dans le MEB tout en prenant en compte ces difficultés. De plus,comme la reconstruction est obtenue par auto-étalonnage de la caméra, l’utilisation des mires n’estplus requise. La sortie finale des techniques présentées est un nuage de points dense, pouvant donccontenir des millions de points, correspondant à la surface de l’objet. / The goal of this work is to obtain a 3D model of an object from its multiple views acquired withScanning Electron Microscope (SEM). For this, the technique of 3D reconstruction is used which isa well known application of computer vision. However, due to the specificities of image formation inSEM, and in microscale in general, the existing techniques are not applicable to the SEM images. Themain reasons for that are the parallel projection and the problems of SEM calibration as a camera.As a result, in this work we developed a new algorithm allowing to achieve 3D reconstruction in SEMwhile taking into account these issues. Moreover, as the reconstruction is obtained through cameraautocalibration, there is no need in calibration object. The final output of the presented techniques isa dense point cloud corresponding to the surface of the object that may contain millions of points.

Cette thèse traite de la reconstruction tridimensionnelle d’une scène rigide à partir d’une collection de photographies numériques, dites vues. Le problème traité est connu sous le nom du "calcul de la structure et du mouvement" (structure-and/from-motion) qui consiste à "expliquer" des trajectoires de points dits d’intérêt au sein de la collection de vues par un certain mouvement de l’appareil (dont sa trajectoire) et des caractéristiques géométriques tridimensionnelles de la scène. Dans ce travail, nous proposons les fondements théoriques pour étendre certaines méthodes de calcul de la structure et du mouvement afin d’intégrer comme données d’entrée, des points d’intérêt réels et des points d’intérêt complexes, et plus précisément des images de points cycliques. Pour tout plan projectif, les points cycliques forment une paire de points complexes conjugués qui, par leur invariance par les similitudes planes, munissent le plan projectif d’une structure euclidienne. Nous introduisons la notion de marqueurs cycliques qui sont des marqueurs plans permettant de calculer sans ambiguïté les images des points cycliques de leur plan de support dans toute vue. Une propriété de ces marqueurs, en plus d’être très "riches" en information euclidienne, est que leurs images peuvent être appariées même si les marqueurs sont disposés arbitrairement sur des plans parallèles, grâce à l’invariance des points cycliques. Nous montrons comment utiliser cette propriété dans le calcul projectif de la structure et du mouvement via une technique matricielle de réduction de rang, dite de factorisation, de la matrice des données correspondant aux images de points réels, complexes et/ou cycliques. Un sous-problème critique abordé dans le calcul de la structure et du mouvement est celui de l’auto-calibrage de l’appareil, problème consistant à transformer un calcul projectif en un calcul euclidien. Nous expliquons comment utiliser l’information euclidienne fournie par les images des points cycliques dans l’algorithme d’auto-calibrage opérant dans l’espace projectif dual et fondé sur des équations linéaires. L’ensemble de ces contributions est finalement utilisé pour une application de suivi automatique de caméra utilisant des marqueurs formés par des couronnes concentriques (appelés CCTags), où il s’agit de calculer le mouvement tridimensionnel de la caméra dans la scène à partir d’une séquence vidéo. Ce type d’application est généralement utilisé dans l’industrie du cinéma ou de la télévision afin de produire des effets spéciaux. Le suivi de caméra proposé dans ce travail a été conçu pour proposer le meilleur compromis possible entre flexibilité d’utilisation et précision des résultats obtenus. / The thesis deals with the problem of 3D reconstruction of a rigid scene from a collection of views acquired by a digital camera. The problem addressed, referred as the Structure-from-Motion (SfM) problem, consists in computing the camera motion (including its trajectory) and the 3D characteristics of the scene based on 2D trajectories of imaged features through the collection. We propose theoretical foundations to extend some SfM paradigms in order to integrate real as well as complex imaged features as input data, and more especially imaged circular points. Circular points of a projective plane consist in a complex conjugate point-pair which is fixed under plane similarity ; thus endowing the plane with an Euclidean structure. We introduce the notion of circular markers which are planar markers that allows to compute, without any ambiguity, imaged circular points of their supporting plane in all views. Aside from providing a very “rich” Euclidean information, such features can be matched even if they are arbitrarily positioned on parallel planes thanks to their invariance under plane similarity ; thus increasing their visibility compared to natural features. We show how to benefit from this geometric property in solving the projective SfM problem via a rank-reduction technique, referred to as projective factorization, of the matrix whose entries are images of real, complex and/or circular features. One of the critical issues in such a SfM paradigm is the self-calibration problem, which consists in updating a projective reconstruction into an euclidean one. We explain how to use the euclidean information provided by imaged circular points in the self-calibration algorithm operating in the dual projective space and relying on linear equations. All these contributions are finally used in an automatic camera tracking application relying on markers made up of concentric circles (called C2Tags). The problem consists in computing the 3D camera motion based on a video sequence. This kind of application is generally used in the cinema or TV industry to create special effects. The camera tracking proposed in this work in designed in order to provide the best compromise between flexibility of use and accuracy.

Cette thèse traite de la reconstruction 3D par stéréophotométrie, qui consiste à utiliser plusieurs photographies d'une scène prises sous le même angle, mais sous différents éclairages. Nous nous intéressons dans un premier temps à des techniques robustes pour l'estimation des normales à la surface, et pour leur intégration en une carte de profondeur. Nous étudions ensuite deux situations où le problème est mal posé : lorsque les éclairages sont inconnus, ou lorsque seuls deux éclairages sont utilisés. La troisième partie est consacrée à l'étude de modèles plus réalistes, à la fois en ce qui concerne les éclairages et la réflectance de la surface. Ces trois premières parties nous amènent aux limites de la formulation classique de la stéréophotométrie : nous introduisons finalement, dans la partie 4, une reformulation variationnelle et différentielle du problème qui permet de dépasser ces limites. / This thesis tackles the photometric stereo problem, a 3D-reconstruction technique consisting in taking several pictures of a scene under different lightings. We first focus on robust techniques for estimating the normals to the surface, and for integrating these normals into a depth map. Then, we study two situations where the problem is ill-posed: when lightings are unknown and when only two images are used. Part 3 is devoted to more realistic models, in terms of lightings and of surface reflectance. These first three parts bring us to the limits of the usual formulation of photometric stereo: we eventually introduce in Part 4 a variational and differential reformulation of this problem which allows us to overcome these limits.

La reconstruction 3D consiste à acquérir des images d’un objet, et à s’en servir pour en estimer un modèle 3D. Dans ce manuscrit, nous développons une méthode de reconstruction basée sur la modélisation par squelette. Cette méthode a l’avantage de renvoyer un modèle 3D qui est un objet virtuel complet (i.e. fermé) et aisément éditable, grâce à la structure du squelette. Enfin, l’objet acquis n’a pas besoin d’être texturé, et entre 3 et 5 images sont suffisantes pour la reconstruction. Dans une première partie, nous étudions les aspects 2D de l’étude. En effet, l’estimation d’un squelette 3D nécessite d’étudier la formation de la silhouette de l’objet à partir de son squelette, et donc les propriétés de sa projection perspective, appelée squelette perspectif. Cette étude est suivie par notre première contribution : un algorithme d’estimation de la projection perspective d’un squelette 3D curviligne, constitué d’un ensemble de courbes. Cet algorithme a toutefois tendance, comme beaucoup d’algorithmes estimant un squelette, à générer des branches peu informatives, notamment sur une image rastérisée. Notre seconde contribution est donc un algorithme d’estimation de squelette 2D, capable de prendre en compte la discrétisation du contour de la forme 2D, et d’éviter ces branches peu informatives. Cet algorithme, d’abord conçu pour estimer un squelette classique, est ensuite généralisé à l’estimation d’un squelette perspectif. Dans une seconde partie, nous estimons le squelette 3D d’un objet à partir de ses projections. Tout d’abord, nous supposons que le squelette de l’objet 3D à reconstruire est curviligne. Ainsi, chaque squelette perspectif estimé correspond à la projection du squelette 3D de l’objet, sous différents points de vue. La topologie du squelette étant affectée par la projection, nous proposons notre troisième contribution, l’estimation de la topologie du squelette 3D à partir de l’ensemble de ses projections. Une fois celle-ci estimée, la projection d’une branche 3D du squelette est identifiée sur chaque image, i.e. sur chacun des squelettes perspectifs. Avec cette identification, nous pouvons trianguler les branches du squelette 3D, ce qui constitue notre quatrième contribution : nous sommes donc en mesure d’estimer un squelette curviligne associé à un ensemble d’images d’un objet. Toutefois, les squelettes 3D ne sont pas tous constitués d’un ensemble de courbes : certains d’entre eux possèdent aussi des parties surfaciques. Notre dernière contribution, pour reconstruire des squelettes 3D surfaciques, est une nouvelle approche pour l’estimation d’un squelette 3D à partir d’images : son principe est de faire grandir le squelette 3D, sous les contraintes données par les images de l’objet. / The principle of 3D reconstruction is to acquire one or more images of an object, and to use it to estimate a 3D model of the object. In this manuscript, we develop a reconstruction method based on a particular model, the skeleton. The main advantages of our reconstruction approach are: we do reconstruct a whole, complete objet, and thanks to the skeleton structure, easily editable. Moreover, the method we propose allows us to free ourselves from constraints related to more classical reconstruction methods: the reconstructed object does not need to be textured, and between 3 and 5 images are sufficient to perform the reconstruction. In the first part, we focus on the 2D aspects of the work. Indeed, before estimating a 3D skeleton, we study the perspective silhouette of the object, and thus evaluate the properties of the perspective projection of the skeleton. Thus, our first contribution is an algorithm estimating the perspective projection of a curvilinear 3D skeleton, consisting of a set of curves. This algorithm, however, like most skeletonisation algorithms, tends to generate non-informative branches, in particular on a rasterized image. Our second contribution is thus an original 2D skeleton estimation algorithm, able to take into account the noise on the contour of the 2D shape, and to avoid uninformative skeleton branches. This algorithm, first designed to estimate a classical skeleton, is then generalized for computing a perspective skeleton. In a second part, we estimate the 3D skeleton of an object from its projections. First, we assume that the skeleton of the considered object is curvilinear. Thus, each estimated perspective skeleton corresponds to the projection of the 3D skeleton, from several viewpoints. The topology of the skeleton is however affected by the perspective projection, so we propose our third contribution: the estimation of the topology of the 3D skeleton based on its projections. Once this topology is estimated, for any 3D branch of the skeleton we indentify its projections on each image, that is a branch on each of the perspective skeletons. From this identification, we triangulate the branches of the 3D skeleton, which is our fourth contribution. Thus, we are able to estimate a curvilinear skeleton associated with a set of images of a 3D object. However, 3D skeletons are not necessarily made up of a set of curves, and some of them also have surface parts. Our last contribution is a new approach for the estimation of a general 3D skeleton (with surface parts) from images, which principle is to increase the 3D skeleton under the constraints given by the different images of the object.