Bachelor and Master Theses

We permanently offer proposals for bachelor and master thesis projects in all areas across our research activities (see our research areas page) and related subjects which cover most topics in Virtual Reality and Scientific Visualization. The thesis topics are usually specified in cooperation with one of our research assistants and/or Prof. Kuhlen taking into account the student's individual interests and his/her previous knowledge as well as the current research agenda of the Virtual Reality group (e.g. in terms of ongoing academic or industrial cooperations). So if you are interested in a thesis project in Virtual Reality, please contact us. In order to guarantee a successful completion of the thesis, we usually expect our student to have

  • taken the "Basic Techniques in Computer Graphics" lecture if you are a bachelor student
  • taken the “Virtual Reality” lecture if you are a master student
  • a good working knowledge of C++
  • or an equivalent qualification.
Below you find a (non-complete) list of currently open theses and the respective supervisors to contact.


Bachelor Thesis: The grid processing library

Scalar, vector, tensor and higher-order fields are commonly used to represent scientific data in various disciplines including geology, physics and medicine. Although standards for storage of such data exists (e.g. HDF5), almost every application has its custom in-memory format. The core idea of this engineering-oriented work is to develop a library to standardize in-memory representation of such fields, and providing functionality for (parallel) per-cell and per-region operations on them (e.g. computation of gradients/Jacobian/Hessian).

Contact:
Ali Can Demiralp, M. Sc.


Bachelor Thesis: Scalar and vector field compression for GPUs based on ASTC texture compression

Scalar and vector fields are N-dimensional, potentially non-regular, grids commonly used to store scientific data. Adaptive Scalable Texture Compression (ASTC) is a lossy block-based texture compression method, which covers the features of all texture compression approaches to date and more. The limited memory space of GPUs pose a challenge to interactive compute and visualization on large datasets. The core idea of this work is to explore the potential uses of ASTC for compression of large 2D/3D scalar and vector fields, attempting the minimize and bound the errors introduced by lossiness.

Contact:
Ali Can Demiralp, M. Sc.


Master Thesis: The multi-device ray tracing library

There are various solutions for ray tracing on CPUs and GPUs today: Intel Embree for shared parallelism on the CPU, Intel Ospray for distributed parallelism on the CPU, NVIDIA OptiX for shared and distributed parallelism on the GPU. Each of these libraries have their pros and cons. Intel Ospray scales to distributed settings for large data visualization, however is bound by the performance of the CPU which is subpar to the GPU for the embarassingly-parallel problem of ray tracing. NVIDIA OptiX provides a powerful programmable pipeline similar to OpenGL but is bound by the memory limitations of the GPU. The core idea of this engineering-oriented work is to develop a library (a) enabling development of ray tracing algorithms without explicit knowledge of the device the algorithm will run on, (b) bringing ease-of-use of Intel Ospray and functional programming concepts of NVIDIA OptiX together.

Contact:
Ali Can Demiralp, M. Sc.


Master Thesis: The multi-device data- and task-parallelism library

There are various solutions for shared and distributed parallelism on CPUs and GPUs today: OpenMP and Intel Threading Building Blocks (TBB) for shared parallelism on the CPU, OpenCL and NVIDIA CUDA for shared parallelism on the GPU, MPI for distributed parallelism on both the CPU and the GPU. The majority of algorithms in scientific data analysis and visualization may be divided into two groups: data-parallel and task-parallel. Data-parallel approaches divide the domain into smaller chunks which are distributed to the processes, whereas task-parallel approaches divide the problem into smaller chunks and distribute them instead. The core idea of this engineering-oriented work is to develop a library enabling development of data-parallel and task-parallel visualization algorithms without explicit knowledge of the device the algorithm will run on.

Contact:
Ali Can Demiralp, M. Sc.


Master Thesis: Numerical relativity library

Numerical relativity is one of the branches of general relativity that uses numerical methods to analyze problems. The primary goal of numerical relativity is to study spacetimes whose exact form is not known. Within this context the geodesic equation generalizes the notion of a straight line to curved spacetime. The core idea of this work is to develop a library for solving the geodesic equation, which in turn enables 4-dimensional spacetime ray tracing. The implementation should at least provide the Schwarzschild and Kerr solutions to the Einstein Field Equations, providing visualizations of non-rotating and rotating uncharged black holes.

Contact:
Ali Can Demiralp, M. Sc.


Master Thesis: Mean curvature flow for truncated spherical harmonics expansions

Curvature flows produce successively smoother approximations of a given piece of geometry, by reducing a fairing energy. Within this context, mean curvature flow is a curvature flow defined for hypersurfaces in a Riemannian manifold (e.g. smooth 3D surfaces in Euclidean space), which emphasizes regions of higher frequency and converges to a sphere. Truncated spherical harmonics expansions are commonly used to represent scientific data as well as arbitrary geometric shapes. The core idea of this work is to establish the mathematical concept of mean curvature flow within the spherical harmonics basis, which is empirically done through interpolation of the harmonic coefficients to the coefficient 0,0.

Contact:
Ali Can Demiralp, M. Sc.


Master Thesis: Orientation distribution function topology

Topological data analysis methods have been applied extensively to scalar and vector fields for revealing features such as critical and saddle points. There is recent effort on generalizing these approaches to tensor fields, although limited to 2D. Orientation distribution functions, which are the spherical analogue to a tensor, are often represented using truncated spherical harmonics expansions and are commonly used in visualization of medical and chemistry datasets. The core idea of this work is to establish the mathematical framework for extraction of topological skeletons from an orientation distribution function field.

Contact:
Ali Can Demiralp, M. Sc.


Master Thesis: Variational inference tractography

Tractography is a method for estimation of nerve tracts from discrete brain data, often obtained through Magnetic Resonance Imaging. The family of Markov Chain Monte Carlo (MCMC) methods form the current standard to (global) tractography, and have been extensively researched to date. Yet, Variational Inference (VI) methods originating in Machine Learning provide a quicker alternative to statistical inference. Stein Variational Gradient Descent (SVGD) is one such method which not only extracts minima/maxima but is able to estimate the complete distribution. The core idea of this work is to apply SVGD to tractography, working with both Magnetic Resonance and 3D-Polarized Light Imaging data.

Contact:
Ali Can Demiralp, M. Sc.


Master Thesis: Block connectivity matrices

Connectivity matrices are square matrices for describing structural and functional connections between distinct brain regions. Traditionally, connectivity matrices are computed for segmented brain data, describing the connectivity e.g. among Brodmann areas in order to provide context to the neuroscientist. The core idea in this work is to take an alternative approach, dividing the data into a regular grid and computing the connectivity between each block, in a hierarchical manner. The presentation of such data as a matrix is non-trivial, since the blocks are in 3D and the matrix is bound to 2D, hence it is necessary to (a) reorder the data using space filling curves so that the spatial relationship between the blocks are preserved (b) seek alternative visualization techniques to replace the matrix (e.g. volume rendering).

Contact:
Ali Can Demiralp, M. Sc.


Bachelor Thesis: Lip Sync in Unreal Engine 4

Computer-controlled, embodied, intelligent virtual agents are increasingly often embedded in various applications to enliven the virtual sceneries. Thereby, conversational virtual agents are of prime importance. To this end, adequate mimics and lip sync is required to show realistic and plausible talking characters. The goal of this bachelor thesis is to enable an effective however easy-to-integrate lip sync in our Unreal projects for text-to-speech input as well as recorded speech.

Prerequisites

  • Programming experience with C++ and Unreal Engine 4
  • Knowledge of reinforcement learning is desirable
  • Basic Knowledge of Computer Graphics (i.e., transformations, geometries, lights, cameras, …)
  • Ability to self-reliant prioritize and manage assigned tasks within a given timeframe
  • Open and transparent communication skills to ensure that we are on the same page


Contact:
Jonathan Ehret née Wendt, M.Sc.


Master Thesis: Generating co-verbal Gestures for a Virtual Human using Recurrent Neural Networks

Virtual Humans can be embedded into virtual environments to guide the user through scenes and teach or point out interesting areas. Thereby their behavior has a large influence on the authenticity of the virtual environment and the immersion of a user. One important aspect to this behavior is their movement during speech: co-verbal gestures. The goal of this thesis is to design, develop and test a system to generate authentic co-verbal gestures using RNNs, e.g., Long-Short-Term-Memory (LSTM) networks. Training data for these networks will be provided. The system should become part of an already existing larger software suite to embed believable Virtual Humans into our framework.

Prerequisites: Good programming skills in C++; knowledge of Machine Learning Techniques is desirable

Photo: ©USC Institute for Creative Technologies

Contact:
Jonathan Ehret née Wendt, M.Sc.


Disclaimer Home Visual Computing institute RWTH Aachen University