In this thesis we consider shape and action modeling problems under the perspective of inverse problem theory. Inverse problem theory proposes a mathematical framework for solving model parameter estimation problems. Inverse problems are typically ill-posed, which makes their solution challenging. Regularization theory and Bayesian statistical methods, which are proposed in the context of inverse problem theory, provide suitable methods for dealing with ill-posed problems. Regarding the application of inverse problem theory in shape and action modeling, we first discuss the problem of saliency prediction, considering a model proposed by the coherence theory of attention. According to coherence theory, salience regions emerge via proto-objects which we model using harmonic functions (thin-membranes). We also discuss the modeling of the 3D scene, as it is fundamental for extracting suitable scene features, which guide the generation of proto-objects. The next application we consider is the problem of image fusion. In this context, we propose a variational image fusion framework, based on confidence driven total variation regularization, and we consider its application to the problem of depth image fusion, which is an important step in the dense 3D scene reconstruction pipeline. The third problem we encounter regards action modeling, and in particular the recognition of human actions based on 3D data. Here, we employ a Bayesian nonparametric model to capture the idiosyncratic motions of the different body parts. Recognition is achieved by comparing the motion behaviors of the subject to a dictionary of behaviors for each action, learned by examples collected from other subjects. Next, we consider the 3D modeling of articulated objects from images taken from the web, with application to the 3D modeling of animals. By decomposing the full object in rigid components and by considering different aspects of these components, we model the object up this hierarchy, in order to obtain a 3D model of the entire object. Single view 3D modeling as well as model registration is performed, based on regularization methods. The last problem we consider, is the modeling of 3D specular (non-Lambertian) surfaces from a single image. To solve this challenging problem we propose a Bayesian non-parametric model for estimating the normal field of the surface from its appearance, by identifying the material of the surface. After computing an initial model of the surface, we apply regularization of its normal field considering also a photo-consistency constraint, in order to estimate the final shape of the surface. Finally, we conclude this thesis by summarizing the most significant results and by suggesting future directions regarding the application of inverse problem theory to challenging computer vision problems, as the ones encountered in this work.

Inverse problem theory in shape and action modeling / Ntouskos, Valsamis. - STAMPA. - (2016 Jun 27).

Inverse problem theory in shape and action modeling

NTOUSKOS, VALSAMIS
27/06/2016

Abstract

In this thesis we consider shape and action modeling problems under the perspective of inverse problem theory. Inverse problem theory proposes a mathematical framework for solving model parameter estimation problems. Inverse problems are typically ill-posed, which makes their solution challenging. Regularization theory and Bayesian statistical methods, which are proposed in the context of inverse problem theory, provide suitable methods for dealing with ill-posed problems. Regarding the application of inverse problem theory in shape and action modeling, we first discuss the problem of saliency prediction, considering a model proposed by the coherence theory of attention. According to coherence theory, salience regions emerge via proto-objects which we model using harmonic functions (thin-membranes). We also discuss the modeling of the 3D scene, as it is fundamental for extracting suitable scene features, which guide the generation of proto-objects. The next application we consider is the problem of image fusion. In this context, we propose a variational image fusion framework, based on confidence driven total variation regularization, and we consider its application to the problem of depth image fusion, which is an important step in the dense 3D scene reconstruction pipeline. The third problem we encounter regards action modeling, and in particular the recognition of human actions based on 3D data. Here, we employ a Bayesian nonparametric model to capture the idiosyncratic motions of the different body parts. Recognition is achieved by comparing the motion behaviors of the subject to a dictionary of behaviors for each action, learned by examples collected from other subjects. Next, we consider the 3D modeling of articulated objects from images taken from the web, with application to the 3D modeling of animals. By decomposing the full object in rigid components and by considering different aspects of these components, we model the object up this hierarchy, in order to obtain a 3D model of the entire object. Single view 3D modeling as well as model registration is performed, based on regularization methods. The last problem we consider, is the modeling of 3D specular (non-Lambertian) surfaces from a single image. To solve this challenging problem we propose a Bayesian non-parametric model for estimating the normal field of the surface from its appearance, by identifying the material of the surface. After computing an initial model of the surface, we apply regularization of its normal field considering also a photo-consistency constraint, in order to estimate the final shape of the surface. Finally, we conclude this thesis by summarizing the most significant results and by suggesting future directions regarding the application of inverse problem theory to challenging computer vision problems, as the ones encountered in this work.
27-giu-2016
File allegati a questo prodotto
File Dimensione Formato  
Tesi dottorato Ntouskos

accesso aperto

Note: Ntouskos: Inverse problem theory in shape and action modeling
Tipologia: Tesi di dottorato
Licenza: Creative commons
Dimensione 35.98 MB
Formato Adobe PDF
35.98 MB Adobe PDF

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11573/876043
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact