Dans ce travail, méthodes d'estimation basées sur des images, également connu sous le nom de méthodes directes, sont étudiées qui permettent d'éviter l'extraction de caractéristiques et l'appariement complètement. L'objectif est de produire pose 3D précis et des estimations de la structure. Les fonctions de coût présenté minimiser l'erreur du capteur, car les mesures ne sont pas transformés ou modifiés. Dans la caméra photométrique estimation de la pose, rotation 3D et les paramètres de traduction sont estimées en minimisant une séquence de fonctions de coûts à base d'image, qui sont des non-linéaires en raison de la perspective projection et la distorsion de l'objectif. Dans l'image la structure basée sur le raffinement, d'autre part, de la structure 3D est affinée en utilisant un certain nombre de vues supplémentaires et un coût basé sur l'image métrique. Les principaux domaines d'application dans ce travail sont des reconstitutions d'intérieur, la robotique et la réalité augmentée. L'objectif global du projet est d'améliorer l'image des méthodes d'estimation fondées, et pour produire des méthodes de calcul efficaces qui peuvent être accueillis dans des applications réelles. Les principales questions pour ce travail sont : Qu'est-ce qu'une formulation efficace pour une image 3D basé estimation de la pose et de la structure tâche de raffinement ? Comment organiser calcul afin de permettre une mise en œuvre efficace en temps réel ? Quelles sont les considérations pratiques utilisant l'image des méthodes d'estimation basées sur des applications telles que la réalité augmentée et la reconstruction 3D ? / In this work, image based estimation methods, also known as direct methods, are studied which avoid feature extraction and matching completely. Cost functions use raw pixels as measurements and the goal is to produce precise 3D pose and structure estimates. The cost functions presented minimize the sensor error, because measurements are not transformed or modified. In photometric camera pose estimation, 3D rotation and translation parameters are estimated by minimizing a sequence of image based cost functions, which are non-linear due to perspective projection and lens distortion. In image based structure refinement, on the other hand, 3D structure is refined using a number of additional views and an image based cost metric. Image based estimation methods are usable whenever the Lambertian illumination assumption holds, where 3D points have constant color despite viewing angle. The main application domains in this work are indoor 3D reconstructions, robotics and augmented reality. The overall project goal is to improve image based estimation methods, and to produce computationally efficient methods which can be accomodated into real applications. The main questions for this work are : What is an efficient formulation for an image based 3D pose estimation and structure refinement task ? How to organize computation to enable an efficient real-time implementation ? What are the practical considerations of using image based estimation methods in applications such as augmented reality and 3D reconstruction ?

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.

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.

[pt] Alvo de intenso estudo da visão computacional, a reconstrução densa 3D teve um importante marco com os primeiros sistemas em tempo real a alcançarem precisão milimétrica com uso de câmeras RGBD e GPUs. Entretanto estes métodos não são aplicáveis a dispositivos de menor poder computacional. Tendo a limitação de recursos computacionais como requisito, o objetivo deste trabalho é apresentar um método de odometria visual utilizando câmeras comuns e sem a necessidade de GPU, baseado em técnicas de Structure from Motion (SFM) com features esparsos, utilizando as informações de uma reconstrução densa. A Odometria visual é o processo de estimar a orientação e posição de um agente (um robô, por exemplo), a partir das imagens. Esta dissertação fornece uma comparação entre a precisão da odometria calculada pelo método proposto e pela reconstrução densa utilizando o Kinect Fusion. O resultado desta pesquisa é diretamente aplicável na área de realidade aumentada, tanto pelas informações da odometria que podem ser usadas para definir a posição de uma câmera, como pela reconstrução densa, que pode tratar aspectos como oclusão dos objetos virtuais com reais. / [en] Aim of intense research in the field computational vision, dense 3D reconstruction achieves an important landmark with first methods running in real time with millimetric precision, using RGBD cameras and GPUs. However these methods are not suitable for low computational resources. Having low computational resources as requirement, the goal of this work is to show a method of visual odometry using regular cameras, without using a GPU. The proposed method is based on technics of sparse Structure From Motion (SFM), using data provided by dense 3D reconstruction. Visual odometry is the process of estimating the position and orientation of an agent (a robot, for instance), based on images. This dissertation compares the proposed method with the odometry calculated by Kinect Fusion. Results of this research are applicable in augmented reality. Odometry provided by this work can be used to model a camera and the data from dense 3D reconstruction, can be used to handle occlusion between virtual and real objects.