Topics

Here you find different topics for student research that can be custom tailored to research projects (INF-PM-FPA,INF-PM-FPG,CMS-PRO), Großer Beleg or student theses in a discussion with the supervisor. As a prerequisite students should have attended one or better several of our bachelor or master level courses. Based on their suitability topics can be labeled with RP (research project), BA (Bachelor thesis or Großer Beleg), MA (Master or Diploma thesis)

Theses, Große Belege and Research Projects

• Immersive Technology

  • Immersive ribbon slicing for video labeling^MA, Lennart Woidtke 
    Inspection and labeling of volumetric data is significantly simplified when reduction to 2D slices is possible. For non planar features, generalized slicing with ribbons allow to stick to efficient 2D interaction for labeling. This work builds on a Godot based immersive ribbon slicing approach and extends it to efficient labeling of moving objects in video data by stacking the video slices into a volume.

  • Immersive explanation of sky light effects^BA,MA, Stefan Gumhold
    Rainbows, god rays, moon coronas, and sun halos are the most important sky light effects that are the result of complicated light propagation effects under specific weather constelations. In this topic immersive explanations in mixed or virtual reality shall be developed that illustrate light propagation and allow for interactive exploration of the parameters that influence the sky light effects.

  • Immersive Storytelling for point cloud based scenes^MA,BA,  Tianfang Lin                   Point cloud allows us to recreate environments with remarkable accuracy, capturing the nuances of real-world spaces. This level of realism immerses users in a virtual environment that closely mirrors the physical world, enabling a deeper sense of presence and connection to the narrative. Point cloud-based scenes provide a three-dimensional representation of the environment, enabling users to explore and interact with the virtual space in a natural and intuitive manner. This freedom of movement fosters a sense of agency and empowers users to engage with the story elements from different angles and perspectives. In this topis, the student should design a set of storytelling patterns and a pipeline from point cloud capturing, editing to recording the story and telling the story. This topic can be a research project or master thesis. 

  • Immersive Selection of Points in Room-scale Point Cloud^MA, Tianfang Lin                With 3D point clouds widely being employed in many fields such as architecture, autonomous driving and archaeology, visualization and exploration of large dense point clouds are increasingly available. Users are expecting to select point in VR efficiently, in this topic, the student should design a set of selection strategies and implement efficient way to select points. This topic can be a bachelor thesis, research project or master thesis.

  • Sandbox App Development^BA,MA,RP, Tianfang Lin

    In previous student projects a sandbox augmented with a kinect and a beamer similar to this, was built together with software that allows simple use of the sandbox together with a VR-headset. Interactive exploration of terrain flooding and a multi-player cooperative game have been studied.  The concept of using the sand surface as input device and the projection as feedback or output medium can serve as hardware platform for a large number of further interactive applications like simulation games, landscape / city / 3D modeling, sound design, and a lot more. After inspection of the sandbox students can design their own topics.

  • AR Viewer of Point Cloud  Stream from Multiple Azure Kinect, Tianfang Lin

• Rendering

  • When does Deep Learning Super Sampling help?^RP,BA, Stefan Gumhold
    Deep Learning Super Sampling (DSSL) is an acceleration technique for real-time rendering that is typically combined with real-time raytracing and volume rendering. It is not clear though whether DSSL does also improve less complex rendering and lighting approach. In this topic the potential of DSSL shall be examined with respect to rendering and lighting approaches of decreasing complexity.

  • Cubic Mipmap Interpolation^BA,MA,RP, Stefan Gumhold

    Mipmaps are used on GPUs for minification filtering to avoid aliasing artefacts during rendering textured 3D content on a pixel grid. Interpolation is limited to multi-linear interpolation in texture space and mipmap level. Multi-linear interpolation can be exploited for cubic interpolation but has not been combined with mipmap based filtering. This combination is to be examined in this work by building mipmap hierarchies that are optimal for cubic filtering and a rendering approach for surfaces and volumes that supports multi-scale cubic interpolation.

  • Accelerating Virtual Human Rendering with Tensor Cores^MA,RP, Mario Henze

    Extensive 3D scanning of static and moving people have let to parametric polygonal mesh models that can adapt shape and pose. To use these in real-time applications like video games or mixed reality collaboration several matrix-vector operations need to be evaluated on a per frame basis. Such operations can be accelerated on tensor cores of current GPUs. This topic shall implement a prototype for evaluation and rendering of parametric models purely on the GPU and evaluate the improvement in runtime performance and the change in energy usage.

• Acquisition & Geometry Processing
  • How sweet, I want to have this cat!^MA Stefan Gumhold
    Parameterized models of humans are getting more and more popular in graphics, vision and entertainmen. Less often considered are parameterized models of animals like SMAL. This can represent different quadruped animals like cats or dogs. It supports shape and pose adaptation but does not provide texture information necessary for a realistic appearance. SMAL comes with a fitting procedure that allows to adapt the pose and shape parameters to a given image. In this topic this feature shall be exploited to extract an appearance texture from a video of an animal - maybe a cat? The texture can either contain just diffuse reflection or for increased realism Gaussian Splats. 
  • Please don't cut me off!^BA,MA Zihan Zhang
    Leaf area is an important criterion to judge the health of plants. This can be easily measured for a leaf by cutting it off, flatening it and scanning it with a 2D scanner. After cutting there is though no way to analyze how leafs grow. In this topic a non invasive leaf acquisition shall be developed based on 3D acquisition of the leaf with the help of rgb cameras. A back light shall be used to allow high contrast in the venetion structure of the leaf that will help tracking over time. As a processing application the acquired 3D shape of the leaf shall be analyzed with respect to surface area and flattend to support comparison to a physical flattening.
  • Malerweg 360°^BA,RP,MA,  Stefan Gumhold (subtopics taken be 3 students)
    Malerweg is one of the most famous hiking trails in Saxon Switzerland split into 8 stages. In this topic the Malerweg should be hiked with a 360° camera in order to acquire a 360° video. This video should then first be georeferenced such that map data can be related to the 360° video. Also partial acquisition of Malerweg is possible. Further tasks for this or further topics:
    • Anonymization by removal of people and number plates from the 360° video
    • 3D reconstruction and immersive exploration of the scenary for example with conversion to a 3D Gaussian Splat scene. (taken)
    • Localization of map data in the video content to extract monument walk bys, 360° views at lookout points, etc.
    • Merging of multiple acquisitions of the same stage at different day times, seasons or weather conditions
    • your own topic
  • Heavy Rain Flooding Risk Quantification^MA,  Stefan Gumhold
    Sealing of soil increases the risk of floodings in case of heavy rain fall that is much more frequent nowadays due to climate change. In addition to soil sealing the geometric structure of the terrain can significantly increase flooding risk. In this thesis simple water balance models shall be used to come up with geometry analysis techniques on digital elevation models to predict flooding risc due to heavy rain fall events. The technique is to be evaluated on known flooding events in the adsence of rivers and to be used to predict risky regions in Germany or Saxony where unsealing measures are important to reduce flooding risk.
• Scientific Visualization 

  • Shadow Visualization in Urban Environments Using Open Data, Lennart Woidtke
    Shadows significantly influence urban planning and the comfort of city dwellers, especially on hot days. This topic leverages detailed open data from Dresden, including information on buildings and trees, to create a realistic 3D visualization of the city, dynamically showcasing available shade at different times of the day and year. Students will use geographic information system (GIS) data to model buildings and individual trees, considering attributes such as tree height, type, crown, and trunk diameters to accurately simulate the shadows they cast. A unique feature of this visualization will be the ability to adjust the shadow coverage based on the size of pedestrians, providing insights into sun exposure for individuals of different heights. The topic will require working with GIS data and implementing a realistic rendering pipeline, with the choice of framework left to the student.

  • In Tube-Visualization, David Groß
    Multivariate data along trajectories has been successfully visualized on tubes in OnTubeVis. For applications like distant heating flow data is often available inside of tube-like structures. This topic explores a variant of the OnTubeVis approach for showing flow attributes inside of tubular structures.
  • Event Evolution Visualization, Stefan Gumhold
    Events arise in very different contexts: traffic events like accidents or traffic jams, particle collisions in particle accelerators, or even discrete steps in the orbit computation of fractals. Showing events in spatial domain allows to efficiently perceive event density. Often also the temporal sequence or evolution of events is of interest.  In this topic visualization techniques shall be developed that facilitate exploration of the time evolution of events. 
  • Particle tracking velocimetry in foams using radiography, Sascha Heitkam (pdf)
• Artificial Intelligence

  • How to make LLMs code explainable action policies?^MA, Diplom, Oleh Shkalikov

    Deep learning-based RL policies lack interpretability even on very simple tasks. Code generation features of current LLMs allow to produce programmatic policies that are inherently interpretable. Reaching state-of-the-art policy performance is challenging though. In this topic LLMs shall be adapted with evolution approaches like AlphaEvolve, that are guided by reinfocement losses. RL benchmarks such as classic control tasks or MuJoCo shall be used to compare performance of the new method compared to traditional deep RL methods. Optionally, an extension to environments with image observations can be examined with vision-language models or zero-shot segmentation models.

  • Universal feature extractor for Atari games, Oleh Shkalikov

    Atari games are one of the most established benchmarks for reinforcement 
    learning with image-based observations. In typical deep RL setups, 
    policies for Atari environments rely on a feature extractor, usually a 
    convolutional network, followed by a policy network and/or a value 
    network. The role of the feature extractor is to convert raw frames into 
    representations that are informative for decision making. However, the 
    standard practice is that each environment trains its own feature 
    extractor from scratch, even though all Atari games share the same input 
    format: four stacked grayscale frames of size 84 by 84.

    The goal of this project is to explore whether it is possible to train a 
    single, universal feature extractor that works across all Atari games. 
    The student will investigate different architectural choices and 
    training approaches, for example VAE, VQ-VAEs or discrete VAEs. Another 
    key question is where and how to train the feature extractor, whether 
    independently from the policy and value networks or jointly with them. 
    The project will examine whether such a unified representation is 
    beneficial for downstream RL performance and whether it can reduce 
    training time or improve generalization and explainability across games.