Template-based Monocular 3-D Shape Reconstruction And Tracking Using Laplacian Meshes

Tien Dat Ngo
EPFL, 2016
Thèse EPFL

Résumé

This thesis addresses the problem of recovering the 3-D shape of a deformable object in single images, or image sequences acquired by a monocular video camera, given that a 3-D template shape and a template image of the object are available. While being a very challenging problem in computer vision, being able to reconstruct and track 3-D deformable objects in videos allows us to develop many potential applications ranging from sports and entertainments to engineering and medical imaging. This thesis extends the scope of deformable object modeling to real-world applications of fully 3-D modeling of deformable objects from video streams with a number of contributions. We show that by extending the Laplacian formalism, which was first introduced in the Graphics community to regularize 3-D meshes, we can turn the monocular 3-D shape reconstruction of a deformable object given correspondences with a reference image into a much better-posed problem with far fewer degrees of freedom than the original one. This has proved key to achieving real-time performance while preserving both sufficient flexibility and robustness. Our real-time 3-D reconstruction and tracking system of deformable objects can very quickly reject outlier correspondences and accurately reconstruct the object shape in 3D. Frame-to-frame tracking is exploited to track the object under difficult settings such as large deformations, occlusions, illumination changes, and motion blur. We present an approach to solving the problem of dense image registration and 3-D shape reconstruction of deformable objects in the presence of occlusions and minimal texture. A main ingredient is the pixel-wise relevancy score that we use to weigh the influence of the image information from a pixel in the image energy cost function. A careful design of the framework is essential for obtaining state-of-the-art results in recovering 3-D deformations of both well- and poorly-textured objects in the presence of occlusions. We study the problem of reconstructing 3-D deformable objects interacting with rigid ones. Imposing real physical constraints allows us to model the interactions of objects in the real world more accurately and more realistically. In particular, we study the problem of a ball colliding with a bat observed by high speed cameras. We provide quantitative measurements of the impact that are compared with simulation-based methods to evaluate which simulation predictions most accurately describe a physical quantity of interest and to improve the models. Based on the diffuse property of the tracked deformable object, we propose a method to estimate the environment irradiance map represented by a set of low frequency spherical harmonics. The obtained irradiance map can be used to realistically illuminate 2-D and 3-D virtual contents in the context of augmented reality on deformable objects. The results compare favorably with baseline methods. In collaboration with Disney Research, we develop an augmented reality coloring book application that runs in real-time on mobile devices. The app allows the children to see the coloring work by showing animated characters with texture lifted from their colors on the drawing. Deformations of the book page are explicitly modeled by our 3-D tracking and reconstruction method. As a result, accurate color information is extracted to synthesize the character's texture.

Official source

https://infoscience.epfl.ch/record/217899?ln=fr

À propos de ce résultat

Cette page est générée automatiquement et peut contenir des informations qui ne sont pas correctes, complètes, à jour ou pertinentes par rapport à votre recherche. Il en va de même pour toutes les autres pages de ce site. Veillez à vérifier les informations auprès des sources officielles de l'EPFL.

Concepts associés

Chargement

Publications associées

Chargement

Tien Dat Ngo
EPFL, 2016
Thèse EPFL

Résumé

Official source

https://infoscience.epfl.ch/record/217899?ln=fr

À propos de ce résultat

Concepts associés (30)

Concepts associés

Chargement

Publications associées (29)

Publications associées

Chargement

Representing and reconstructing 3D deformable shapes are two tightly linked problems that have long been studied within the computer vision field. Deformable shapes are truly ubiquitous in the real world, whether be it specific object classes such as humans, garments and animals or more abstract ones such as generic materials deforming under an external force. Practical computer vision algorithms must be able to understand the shapes of objects in the observed scenes to unlock the wide spectrum of much sought after applications ranging from virtual try-on to automated surgeries.Automatic shape reconstruction is known to be an ill-posed problem, especially in the common scenario of a single image input. Therefore, the modern approaches rely on deep learning paradigm which has proven to be extremely effective even for the severely under-constrained computer vision problems. We, too, exploit the success of data-driven approaches, however, we also show that generic deep learning models can greatly benefit from being combined with explicit knowledge originating in computational geometry. We analyze the use of various 3D shape representations and we distinctly focus on one of them, the atlas-based representation, which turns out to be especially suitable for modeling deformable shapes and which we further improve and extend.The atlas-based representation models the surfaces as an ensemble of continuous functions and thus allows for arbitrary resolution and analytical surface analysis. We identify its major shortcomings, namely the patch collapse, patch overlap and strong mapping distortions, and we propose novel regularizers based on analytically computed properties of the reconstructed surfaces. Our approach counteracts the aforementioned drawbacks while yielding higher reconstruction accuracy.We dive into the problematics of atlas-based shape representation deeper and focus on another design flaw, the global inconsistency of the mappings. While it is not reflected in quantitative metrics, it is detrimental to the visual quality of the reconstructed surfaces. Specifically, we design loss functions encouraging intercommunication among the mappings which pushes the resulting surface towards a C1 smooth function and thus dramatically improves the visual quality.Furthermore, we adapt the atlas-based representation so that it could model a full sequence of a deforming object in a temporally-consistent way. The goal is to produce such reconstruction where each surface point always represents the same semantic point on the target GT surface. To achieve such behavior, we note that if each surface point deforms close-to-isometrically, its semantic location likely remains unchanged. Practically, we make use of the Riemannian metric, and force it to remain point-wise constant throughout the sequence. The experiments show that our method yields SotA results on correspondence estimation task.Finally, we look into a particular problem of monocular texture-less deformable shape reconstruction. We propose a multi-task learning approach which jointly produces a normal map, a depth map and a mesh corresponding to the observed surface. We show that producing multiple different 3D representations of the same objects results in higher reconstruction quality. We acquire a large real-world annotated dataset of texture-less deforming objects and we release it for public use.

EPFL2022

Chargement

In this thesis, we focus on the problem of recovering 3D shapes of deformable surfaces from a single camera. This problem is known to be ill-posed as for a given 2D input image there exist many 3D shapes that give visually identical projections. We present three methods which make headway towards resolving these ambiguities. We believe that our work represents a significant step towards making surface reconstruction methods of practical use. First, we propose a surface reconstruction method that overcomes the limitations of the state-of-the-art template-based and non-rigid structure from motion methods. We neither track points over many frames, nor require a sophisticated deformation model, or depend on a reference image. In our method, we establish correspondences between pairs of frames in which the shape is different and unknown. We then estimate homographies between corresponding local planar patches in both images. These yield approximate 3D reconstructions of points within each patch up to a scale factor. Since we consider overlapping patches, we can enforce them to be consistent over the whole surface. Finally, a local deformation model is used to fit a triangulated mesh to the 3D point cloud, which makes the reconstruction robust to both noise and outliers in the image data. Second, we propose a novel approach to recovering the 3D shape of a deformable surface from a monocular input by taking advantage of shading information in more generic contexts than conventional Shape-from-Shading (SfS) methods. This includes surfaces that may be fully or partially textured and lit by arbitrarily many light sources. To this end, given a lighting model, we learn the relationship between a shading pattern and the corresponding local surface shape. At run time, we first use this knowledge to recover the shape of surface patches and then enforce spatial consistency between the patches to produce a global 3D shape. Instead of treating texture as noise as in many SfS approaches, we exploit it as an additional source of information. We validate our approach quantitatively and qualitatively using both synthetic and real data. Third, we introduce a constrained latent variable model that inherently accounts for geometric constraints such as inextensibility defined on the mesh model. To this end, we learn a non-linear mapping from the latent space to the output space, which corresponds to vertex positions of a mesh model, such that the generated outputs comply with equality and inequality constraints expressed in terms of the problem variables. Since its output is encouraged to satisfy such constraints inherently, using our model removes the need for computationally expensive methods that enforce these constraints at run time. In addition, our approach is completely generic and could be used in many other different contexts as well, such as image classification to impose separation of the classes, and articulated tracking to constrain the space of possible poses.

EPFL2012

Chargement

Dense deformation field estimation for atlas registration using the active contour framework

Valérie Duay

A key research area in computer vision is image segmentation. Image segmentation aims at extracting objects of interest in images or video sequences. These objects contain relevant information for a given application. For example, a video surveillance application generally requires to extract moving objects (vehicles, persons or animals) from a sequence of images in order to check that their path stays conformed to the regulation rules set for the observed scene. Image segmentation is not an easy task. In many applications, the contours of the objects of interest are difficult to delineate, even manually. The problems linked to segmentation are often due to low contrast, fuzzy contours or too similar intensities with adjacent objects. In some cases, the objects to be extracted have no real contours in the image. This kind of objects is called virtual objects. Virtual objects appear especially in medical applications. To draw them, medical experts usually estimate their position from surrounding objects. The problems related to image segmentation can be greatly simplified with information known in advance on the objects to be extracted (the prior knowledge). A widely used method consists to extract the needed prior knowledge from a reference image often called atlas. The goal of the atlas is to describe the image to be segmented like a map would describe the components of a geographical area. An atlas can contain three types of information on each object being part of the image: an estimation of its position in the image, a description of its shape and texture, and the features of its adjacent objects. The atlas-based segmentation method is rather used when the atlas can characterize a range of images. This method is thus especially adapted to medical images due to the existing consistency between anatomical structures of same type. There exist two types of atlas: the determinist atlas and the statistical atlas. The determinist atlas is an image which has been selected or computed, to be the most representative of an image category to be segmented. This image is called intensity atlas. The contours of the objects of interest (the objects to be extracted in images of the same type) have been traced manually on the intensity atlas, or by using a semi-automatic method. A label is often attributed to each one of these objects in order to differentiate them. In this way, we obtain a labeled version of the atlas called labeled atlas. The statistical atlas is an atlas created from a database of images in order to be the most representative of a certain type of images to be segmented. In this atlas, the position and the features of the objects of interest depend on statistical measures. In this thesis, we are focused on the use of determinist atlases for image segmentation. The segmentation process with a determinist atlas consists to deform the objects delineated in the atlas in order to better align them with their corresponding objects in the image to be segmented. To perform this task, we have distinguished two types of approaches in the literature. The first approach consists to reduce the segmentation problem in an image registration problem. First of all, a dense deformation field that registers (i.e. puts in point-to-point spatial correspondence) the atlas to the image to be segmented, is explicitly computed. Then, this transformation is used to project the assigned labels onto each atlas structure on the image to be segmented. The advantage of this approach is that the deformation field computed from the registration of visible contours allows to easily estimate the position of virtual objects or objects with fuzzy contours. However, the methods currently used for the atlas registration are often only based on the intensity atlas. That means that they do not exploit the object-based information that can be obtained by combining the intensity atlas with its labeled version. In the second approach, the atlas contours selected by the labeled atlas are directly deformed without using a geometrical deformation. For that, this approach is based on matching contour techniques, generally called deformable models. In this thesis, we are interested to a particular type of deformable models, which are the active contour segmentation models. The advantage of the active contour method is that this segmentation technique has been designed to exploit the image information directly linked to the object to be delineated. By using object-based information, active contour models are frequently able to extract regions where the atlas-based segmentation method by registration fails. On the other hand, the result of this local segmentation method is very sensitive to the initial atlas contour position regarding to the target contours. On the other hand, this local segmentation method is very sensitive to the initial position of the atlas contours: the closer they are to the contours to be detected, the more robust the active contour-based segmentation will be. Besides, this segmentation technique needs prior shape models to be able to estimate the position of virtual objects. The main objective of this thesis is to design an algorithm for atlas-based segmentation which combines the advantages of the dense deformation field computed by the registration algorithms, with local segmentation constraints coming from the active contour framework. This implies to design a model where the registration and segmentation by active contours are jointly performed. The atlas registration algorithm that we propose is based on a formulation allowing the integration of any segmentation or contour regularization forces derived from the theory of the active contours in a non parametric registration process. Our algorithm led us to introduce the concept of hierarchical atlas registration. Its principle is that the registration of the main image objects helps the registration of depending objects. This allows to bring progressively the atlas contours closer to their target and thus, to limit the risk to be stuck in a local minimum. Our model had been designed to be easily adaptable to various types of segmentation problems. At the end of the thesis, we present several examples of atlas registration applications in medical imaging. These applications highlight the integration of manual constraints in an atlas registration process, the modeling of a tumor growth in the atlas, the labelization of the thalamus for a statistical study on neuronal connections, the localization of the subthalamic nucleus (STN) for deep brain stimulation (DBS) and the compensation of intra-operative brain shift for neuronavigation systems.

EPFL2008

Chargement

EPFL2022

EPFL2012

Dense deformation field estimation for atlas registration using the active contour framework

Valérie Duay

EPFL2008