[visionlist] 4 PhD positions at CVC/IMAGINE

[Renaud Marlet]

4 PhD positions are open at CVC/IMAGINE:

(1) Registering textureless images on a 3D model and application to thermal images and building models

The goal of the proposed thesis is to study the registration of images on a 3D model with little or no texture correlation. Registering in this case has to be based on other kinds of features, such as edges and  contours. One application is to register thermal images of a building on an existing 3D building model. In fact, thermal images often consist of isotemperature patches with a more or less uniform color, i.e., basically textureless. Such a registration not only allows moving virtually in a 3D thermal model for diagnosis, but also identifying and finely locating on the model elements that are normally invisible. It could also be used to refine simulations of building energy performance.

Full description, searched profile and application instructions: see
http://imagine.enpc.fr/positions/theses2012/phd-by2012-1-en.pdf

*****

(2) Global modeling of geometry and semantics, and application to the reconstruction of technical equipment

The objective of the proposed thesis is to study a new framework for the automatic interpretation of complex 3D scenes. The hypothesis is that the 3D data can be approximated by a combination of geometric primitives (cylinders, cubes, spheres, etc.) that are constrained by the type of the observed scene, given domain-specific rules.  The idea here is to jointly address the recognition of geometric primitives and their optimal combination in accordance with the semantic composition rules. The originality lies in the global and simultaneous processing of geometric and semantic estimation of the entire scene. Solving this large-scale optimization problem on discrete and continuous variables can borrow some techniques used for shape grammar parsing, including "top-down" statistical inference as well as "bottom-up" detection-based analysis. Expected benefits are an improved treatment of complex scene and a greater flexibility thanks to an easy adaptation of the primitive dictionary to handle new types of scenes or objects. The research will be guided by a case study concerning the reconstruction of the technical equipment in a building.

Full description, searched profile and application instructions: see
http://imagine.enpc.fr/positions/theses2012/phd-by2012-2-en.pdf

*****

(3) Shape grammar learning and application to building model reconstruction

The goal of the proposed thesis is to study the automatic and semi-automatic learning of grammars from examples. This will reduce the need for experts to manually design grammars, and thus help in rapidly constituting a database of grammars capable of handling a wide variety of object types. Learning will address different aspects of a grammar:
- the choice of structural rules, which break down a complex element in more simple sub-elements that are related,
- the statistical distribution of values that parameterize the terminal and non-terminal elements of a grammar (e.g., geometric information, photometric clues),
- the statistical distribution of the relationships between grammatical elements, which can predict the nature and parameters of invisible elements in the data (e.g., roofs).

Full description, searched profile and application instructions: see
http://imagine.enpc.fr/positions/theses2012/phd-cs2102-en.pdf

*****

(4) Higher Order Graph-based 2D-to-3D Multi-modal Image Fusion/Registration

The aim of this thesis is to introduce a 2D-3D registration paradigm that explores higher order graphical models. Higher order interactions between points will be the driving force to determine the observed plane while pair-wise and singleton terms will produce the in-plane deformation. Due to the use of graphical model the proposed solution will be extremely flexible and independent on the nature of images to be fused – examples could refer to a navigation setting either in the context of military settings or computer assisted surgery, and could produce real-time performance if appropriate use of parallel programming architectures is considered. The scientific impact of the thesis could be tremendous as well since the combination of low-rank and high-rank graphical models is an open research topic that requires research both in terms of optimization as well as in terms of dimensionality reduction. Such models could be used to cope with the most challenging problems in computer vision, like pose invariant object detection/ recognition, image segmentation, 3D reconstruction, etc.

Full description, searched profile and application instructions: see
http://imagine.enpc.fr/positions/theses2012/phd-th2012-en.pdf