Search

Geometric approximation of urban objects with compact and accurate representation is a challenging problem that concerns both computer vision and computer graphics communities. Existing literature mainly focuses on reconstruction from high-quality point clouds obtained by laser scanning which are too costly for many practical scenarios. This motivates the investigation into the problem of geometric approximation from low-budget image data. Dense reconstruction from a collection of images is made possible by recent advances in multi-view stereo techniques, yet the resulting point cloud is often far from perfect for generating a compact model. In particular, our goal is to describe the captured scene with a compact and accurate representation.In this thesis, we propose two generic algorithms which address different aspects of image-based geometric approximation. First, we present an algorithm for extracting and vectorizing objects in images with polygons. Second, we present a global registration algorithm from multi-modal geometric data, typically 3D point clouds and surface meshes. Both approaches exploit detection of linear geometric primitives to approximate either 2D silhouettes with polygons consisting of line segments, or 3D point sets with a collection of planar shapes. The proposed algorithms could be used sequentially to form a pipeline for geometric approximation of an urban object from a set of image data, consisting of an overhead shot for coarse model extraction and multi-view stereo data for point cloud generation. We demonstrate the robustness and scalability of our methods for structured scenes and objects, as well as applicative potential for free-form objects., L'approximation géométrique d'objets urbains avec une représentation compacte et précise est un problème difficile en vision par ordinateur et en infographie. La littérature existante se concentre principalement sur la reconstruction à partir de nuages de points de haute qualité obtenus par balayage laser qui sont trop coûteux pour de nombreux scénarios pratiques. Ceci motive l'investigation du problème d'approximation géométrique à partir de données image. La reconstruction dense à partir d'une collection d'images est rendue possible par les progrès récents des techniques de stéréoscopie multi-vues, mais le nuage de points résultant est souvent trop imparfait pour créer un modèle compact. En particulier, nous visons à décrire la scène capturée avec une représentation compacte et précise.Dans cette thèse, nous proposons deux algorithmes génériques qui abordent différents aspects de l'approximation géométrique basée image. Dans un premier temps, nous présentons un algorithme d'extraction et de vectorisation d'objets dans des images avec des polygones. Dans un second temps, nous présentons un algorithme de recalage global à partir de données géométriques multimodales, typiquement des nuages de points 3D et des maillages surfaciques. Les deux approches exploitent la détection de primitives géométriques pour approcher soit des formes 2D avec des polygones formés à partir de segments de ligne, soit des ensembles de points 3D avec une collection de formes planes. Les algorithmes proposés peuvent être utilisés de manière séquentielle pour former une chaîne de traitement pour l'approximation géométrique d'un objet urbain à partir d'un ensemble de données d'image, composé d'une prise de vue aérienne pour l'extraction de modèles grossiers et de données stéréo multi-vues pour la génération de nuages de points. Nous démontrons la robustesse et l'évolutivité de nos méthodes pour les scènes et objets structurés, ainsi que le potentiel applicatif pour les objets de forme libre.

International audience; We present a global registration algorithm for multi-modal geometric data, typically 3D point clouds and meshes. Existing feature-based methods and recent deep learning based approaches typically rely upon point-to-point matching strategies that often fail to deliver accurate results from defect-laden data. In contrast, we reason at the scale of planar shapes whose detection from input data offers robustness on a range of defects, from noise to outliers through heterogeneous sampling. The detected planar shapes are projected into an accumulation space from which a rotational alignment is operated. A second step then refines the result with a local continuous optimization which also estimates the scale. We demonstrate the robustness and efficacy of our algorithm on challenging real-world data. In particular, we show that our algorithm competes well against state-of-the-art methods, especially on piece-wise planar objects and scenes.

International audience; Recent years have seen an explosion of the number of camera modules integratedinto individual consumer mobile devices, including configurations that contain multi-ple different types of image sensors. One popular configuration is to combine an RGBcamera for color imaging with a monochrome camera that has improved performancein low-light settings, as well as some sensitivity in the infrared. In this work we in-troduce a method to combine simultaneously captured images from such a two-camerastereo system to generate a high-quality, noise reduced color image. To do so, pixel-to-pixel alignment has to be constructed between the two captured monochrome and colorimages, which however, is prone to artifacts due to parallax. The joint image recon-struction is made robust by introducing a novel artifact-robust optimization formulation.We provide extensive experimental results based on the two-camera configuration of a commercially available cell phone.

Optical Coherence Tomography (OCT) is a coherence-gated, micrometer-resolution imaging technique that focuses a broadband near-infrared laser beam to penetrate into optical scattering media, e.g. biological tissues. The OCT resolution is split into two parts, with the axial resolution defined by half the coherence length, and the depth-dependent lateral resolution determined by the beam geometry, which is well described by a Gaussian beam model. The depth dependence of lateral resolution directly results in the defocusing effect outside the confocal region and restricts current OCT probes to small numerical aperture (NA) at the expense of lateral resolution near the focus. Another limitation on OCT development is the presence of a mixture of speckles due to multiple scatterers within the coherence length, and other random noise. Motivated by the above two challenges, a multiple scattering model based on Rytov approximation and Gaussian beam optics is proposed for the OCT setup. Some previous papers have adopted the first Born approximation with the assumption of small perturbation of the incident field in inhomogeneous media. The Rytov method of the same order with smooth phase perturbation assumption benefits from a wider spatial range of validity. A deconvolution method for solving the inverse problem associated with the first Rytov approximation is developed, significantly reducing the defocusing effect through depth and therefore extending the feasible range of NA.