Die Berechnung des Zusammenhangs zwischen mehreren Bilder ist eine wichtige Teilaufgabe vieler Anwendungen in den Bereichen Computer Vision und Augmented Reality. Ein Beispiel für solche Anwendungen ist die Generierung von Panoramabildern aus mehreren Aufnahmen. Eine Standardlösung für das Problem des Bildmatchings ist das Berechnen von Feature-Punkten auf den Eingabedaten. Aus Matchings zwischen Feature-Punkten verschiedener Bilder lässt sich dann die Beziehung zwischen den Bildern bestimmen. Die bisher gebräuchlichen Verfahren funktionieren bereits sehr zuverlässig, sind aber selbst auf modernen PCs bei weitem nicht schnell genug für Anwendungen in Echtzeit. Dies ist insbesondere im Bereich Augmented Reality ein Problem, da dort die Anwendungen in Echtzeit ablaufen müssen und die Zielplattform hauptsächlich Mobiltelefone und Smartphones sind.
In dieser Bachelorarbeit sollen deshalb verschiedene Teilbausteine eines Feature-Extraktionsalgorithmus so vereinfacht werden, das die resultierenden Algorithmen schnell genug für Mobilgeräte werden. Die Vereinfachung besteht dabei darin die Funktionsweise des ursprünglichen Algorithmus grob zu approximieren. Die Hauptfragestellung ist welche Teile sich wie stark vereinfachen lassen ohne das die Algorithmen ihre Robustheit verlieren. Dazu sollen Bilder miteinander verglichen werden, die aus unterschiedlichen Blickwinkeln und unterschiedlichen Entfernungen aufgenommen wurden oder Störfaktoren wie teilweise Verdeckung oder eine Änderung der Lichtverhältnisse enthalten. Die Fähigkeit der Algorithmen diese Bilder miteinander in Verbindung bringen zu können wird dann mit dem Geschwindigkeitsgewinn der Approximationen verglichen. Zum Vergleich stehen sowohl Desktop PCs als auch Ultra Mobile PCs und ein iPhone zur Verfügung.

Ansprechpartner: Torsten Sattler

An image mosaic in computer graphics is a regular arrangement of so-called "tile images" which seen from a distance suggest another completely, different image: the "image mosaic". We have developed a mosaicing algorithm for creating what we call genuine image mosaics (short GIZMOs), image mosaics composed of source images from a large database without any additional color corrections. We apply differently sophisticated image descriptors to find best matching images. In order to achieve high approximation quality as well as to preserve the mosaicing effect in our results we differentiate between feature and non-feature tiles and apply slightly different image descriptors.

Given the GIZMO algorithm, this theses will combine a number of interesting problems to be addressed and tasks to be solved. Some of them are: A web application should be developed as a frontend for the GIZMO algorithm. How can the image database be preprocessed in order to get rid of unwanted images before the actual mosaic computation? Is there a way to systematically enlarge the database in order to improve the final mosaicing results? Recently introduced color space transformations could be explored in order to improve the matching procedure. Since the GIZMO algorithm strongly relies on the fact that a huge database is used (currently over one million images), it is currently not possible for the user to actually control which images to use in the final mosaic. There should be a way to algorithmically allow for more user control in this particular direction.

 

Contact: Darko Pavic

Mesh repair or mesh fixing is one of the crucial functional parts in the mesh generation pipeline. One of the main tasks in this context is the one of hole filling, where one is interested in recovering the missing parts in 3D models. For the case when 3D models are represented by triangle meshes this task was often addressed in the past and there exist a number of methods which are able to more or less plausibly recover the missing parts.

Although trinagle meshes are still the most often used representation of 3D models in computer graphics, quad meshes are the representation of choice when the structural information is to be captured as good as possible.  Quad meshes are therefore especially preferred in CAD-based modelling or in simulations (finite elements, etc.).

In this thesis a hole filling approach on quad meshes should be implemented. One of the existing approaches on triangle meshes can be used and appropriately adapted in order to be applicable to quad meshes.

 

Contact: Darko Pavic

Watching recent animated movies it is obvious that facial animation and modeling became an important task in the last years. Moreover the synthetic generation of facial expressions is also usefull to support psychological studies. In many of those studies subjects have to do so called emotion recognition tasks. Nowadays the stimuli are not constrained to photos or videos, computer generated facial expressions became more important since they are easy to generate and show a realistic appearance.
One of our current research projects is the development of such a dynamic face model. The goal of this bachelor project is the generation of a generic head model, which is capable to integrate our deformable facial region by deforming the geometry and adapting the skin color of the head model. An integrated stochastic model to synthesize rigid movements of the head will further increase the realistic appearance of the whole model.

 

Contact: Dominik Sibbing

In daily clinic routine and medical research different image modalities like Magnetic Resonance Imaging(MRT), Computer Tomography(CT), Positron Emission Tomography(PET) or Diffusion Tensor Imaging(DTI) became important tools to understand and diagnose the anatomy of humans. Usually the resulting images contain a lot of informations and in most cases only experts can identify anatomical structures and put them in an overall picture.

In this project you will extract the geometry from different image modalities to build an anatomical model from the human brain. This model will contain the structure of the brain itself with white and grey matter, it should show the location of single blood vessels and could contain nerve fibers.

 

Contact: Dominik Sibbing

3D content creation gets more and more important today. For automotive industry 3D models are used to design and optimize cars without having to build them. Another example is game and video content creation where whole characters or scenes are generated in a 3D modeling system.
A common approach for generating 3D models is to reuse and recombine existing parts or surfaces to form a new object. The challenging part is to deform these separate parts and assemble them to a smooth surface without any gaps or seams between them while keeping the user interaction at a minimum.
In this project you will use state of the art algorithms to implement a system creating complex 3D models from a set of simple parts. A large geometry processing architecture called OpenFlipper provides basic functionality such as rendering or selections and will be the basis for the implementation. Finally the system should then be used to create 3D content.

Contact: Jan Möbius

Many important parts of computer graphics and geometry processing rely on the unfolding or flattening of 3D objects. Such techniques are required for example when computing parametrizations of 3D objects or when a 3D surface shall be textured with a 2D image.

Some simple open objects can easily be embedded in a plane, the difficulty arises with closed objects (e.g. a sphere) or objects with higher genus (e.g. a torus). To be able to flatten complex geometries they first need to be cut open so that they contain a single boundary.

Techniques on how to find good cuts already exist, however, as one can imagine, it can be quite difficult to picture what the planar embedding of complex geometries will actually look like when flattened.

The goal of this project is to visualize this process. A system shall be developed which takes an object and a set of cuts as input and then computes an appropriate planar embedding and visualizes the process step by step forming an animation.

The implementation will be based on the OpenFlipper framework which provides basic rendering and interaction functionality.

 

Contact: Henrik Zimmer

CAD models of automotive shapes often consist of a large number of surface patches that contain gaps between them, intersect arbitrarily, and lack a consistent normal orientation. In one of our current research projects we simplify these complex meshes by covering them with a manifold mesh that approximates the basic underlying geometry. The extraction of the basic geometry forms a basis for further mesh processing like deformation or morphing.

The goal of this bachelor thesis is to develop methods that allow for manipulating these cover meshes. Repositioning a vertex should e.g. pull a certain region of the cover mesh over the CAD model thereby assuring that the model's basic geometry is still represented. Special attention should be paid to the development of user interaction functionality that provides intuitive control over the result of the mesh manipulation.

The implementation will be based on the OpenFlipper framework which provides basic rendering and interaction functionality.

 

Contact: Ellen Dekkers

Digital cameras have become commonplace, especially as webcams or integrated in laptops and mobile phones. While the sensor chips of these cameras often yield acceptable quality, the weakest component of simple cameras usually is their lens: Low- and also mid-priced lenses exhibit sometimes strong radial distortion (better known as pincushion or barrel distortion).

In this project we want to develop a software tool that consists of two major components. First, it should be capable of precisely computing the distortion parameters of a camera with respect to a standard distortion model. This involves the detection of a specially prepared pattern from which the distortion parameters can be computed. In addition, the tool is then supposed to remove the distortion of images as batch process for a given set of image files and also as a real-time process for a live camera stream. This is achieved by applying a suitable transformation to the images. Both tasks can efficiently be solved by exploiting the computational power of todays graphics processors (GPUs) and Nvidia's CUDA framework.

Contact: Martin Habbecke

Augmented Reality applications enrich the real world with virtual information. See-through displays, i.e., transparent displays integrated into glasses, are a common way to realize the augmentation by rendering virtual content over the real scene as seen by the user. For such applications it is crucial to have precise knowledge of the current viewing position and orientation of the user -- otherwise a convincing scene augmentation is not possible and the rendered content is rather perceived as disturbing. In addition, when the user is moving, his position has to be updated in real-time to sustain the users experience.

The goal of this bachelor thesis project is the construction of a simple augmented reality system based on a high-framerate digital camera. To simplify the task, we will additionally use sensors for orientation and motion that help keeping track of the cameras position. To achieve the highest performance possible, this project will make use of Nvidia's CUDA framework wherever possible.

Contact: Martin Habbecke

 

In the near term future, most video content will be available in 3D. 3D does not only mean stereo video, but also free viewpoint video, where a user can navigate around and change its viewpoint in a similar way as it is already known from today's video games. Free viewpoint video is recorded using arrays of cameras. Rendering is typically done using 3D or stereo displays. An important factor for users to enjoy free viewpoint video on mobile devices is availability of suitable user interfaces. While design of good user interfaces for free viewpoint selection is an open question as of today, modern mobile phones already have advanced built-in sensors such as video cameras, acceleration sensors, and touch pads, which may help in providing user-friendly viewpoint control mechanisms. With acceleration sensors and touch pads, viewpoint control data may be acquired based on movement of the phone and user interaction on the display, respectively. Video images of the user acquired from the phone's built-in camera, after analysis using head tracking may serve as an alternative or additional control input. The aim of this thesis is to investigate use of advanced input sensors for free viewpoint control. A particular focus is on investigation of head tracking for mobile devices.

 

This topic will be worked on in cooperation with Ericsson, Aachen

 

Contact: Leif Kobbelt