Thesis Topics

Topics for Master and Bachelor Theses

If you are interested in a conducting a thesis project in visual computing, please contact any of the group members to discuss further details. We recommend to first take an advanced course (i.e., beyond the first/second year introductory courses) or seminar with us as a preparation, but this is not a strict requirement if we can find a common topic based on your prior knowledge.

Below, we are listing a few example topics. This list is not exhaustive but meant to give a rough impression of suitable topics. Of course, it would be great if you were interested in one of the specific project ideas listed.

Computer Graphics Topics

  • 3D self-localization and mapping with a dynamic 3D scanner. You are given a dynamic 3D point cloud scanner, such as a Microsoft "Kinect". The device provides a stream of 3D points that are sampled from the environment the scanner currently sees, and the scanner is moved by a human user through a scene (without the system knowing the motion path). The task is now to assemble this stream of 3D points into a consistent scene (mapping), putting each individual scan in the right place (self-localization). Challenges arise due to symmetric geometry (repeating objects) and noise/occlusion and other measurement artifacts.
    The topic is suitable for bachelor (basic pipeline) and master (advanced processing) projects.
  • Example-based generative data models (images + 3D scenes). This is a timely and very interesting topic. Can we learn how classes of 3D objects or 2D images are structured from examples? This means, we just show our algorithm a few examples of what we want (for example, 3D models of cars or castles, or paintings of people) and we want the computer to generalize and create similar content automatically. A variety of techniques exist - from non-parametric texture synthesis (which can be implemented in 10 lines of C++ code and can generalize from a single image) to adversarially trained deep networks (which needs only a few more lines of code in Phython+Tensor Flow, however, along with a million example images). Generally, the 3D case is (seems to be?) more challenging than the 2D case in terms of implementation effort.
    The topic is also suitable for bachelor and master thesis projects.

Computer Vision / Machine Learning Topics

  • 3D object classification in medical CT data. The task is to train a classifier for 3D volume data that recognizes features of medical and/or anatomical relevance in a 3D CT scan. For example, a simple task would be to localize specific bones or organs in a 3D scan. A more complex task would be to recognize medical conditions from example data. Recent advances in computer vision (in particular, representation learning methods such as deep convolutional neural networks) allow us to get quite impressive recognition performance in such tasks.
    This area could be explored in a bachelor thesis (basics) or a master thesis (more complex recognition tasks).
  • 3D object classification in point cloud scans. Similar idea as above, but using point cloud scans from 3D scanners as data source.
  • Improving deep learning methods. Can we use ideas from the computer graphics toolbox (structure models and data representations) in order to improve the learning efficiency of deep neural networks? This would be an advanced master thesis topic for students who are a bit theoretically inclined (not afraid of a bit of math).

Interdisciplinary Research

  • Medical data classification (as discussed above; collaboration with the medical school).
  • Pattern recognition in atmospheric simulations. The goal is to classify and find flow patterns in atmospheric simulation data. This could be done in a supervised (we have examples of what we are looking for) and unsupervised settings (we want to cluster repetitive structures). This project topic would be offered in collaboration with the Institute for Physics of the Atmosphere.
    The topic could be formed into a bachelor or master thesis.
  • Machine learning and deep networks for coarse-graining in multi-scale simulations. The topic says it all - can we learn how to conduct simulations on a very coarse (and easier to compute) level of detail such that the effects on the fine scale are predicted in a qualitatively correct way? There have been some recent, exciting ideas involving deep neural networks proposed in the literature that we could follow up on.
    This topic would be suitable as a master thesis project; a reasonably strong background in physics or mathematics would be highly recommended.

Further topics

Do you have an idea of your own that is related to visual computing? Or are you interested in a specific direction / topic area of that flavor? Do not hesitate to contact any member of our group for a discussion.