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 master level courses.

Theses, Große Belege and Research Projects

  • Brain Twisters (4D and more)
    • Efficient Out-of-Core Slicing of Multi-Dimensional Data, Stefan Gumhold
      Here we consider a scalar dataset organized in an nD-Array or n-Tensor, which is a vector for n=1, a matrix for n=2 and a volume for n=3. Besides space dimensions (x,y,z) further dimensions like time (t) and frequency (f) can be sampled providing multi-dimensional datasets with n>4. In practical applications these are very large and won't fit in main memory. This thesis should investigate an out-of-core strategy to support 1D, 2D and optionally 3D slicing of the data along any combination of axes. There core of this thesis is to come up with a data organization on disk that supports real-time streaming of 1D, 2D, 3D slices to support interactive exploration of the multi-dimensional dataset.
  • Immersive Technology

    • Rendering Volumetric Depth Images in Virtual Reality Aryaman Gupta
      Volumetric Depth Images (VDIs) are compact view-dependent representations of volume data that can be rendered much faster than the original data. They are generated by casting rays through the volume and decomposing the volume into ray-segments that store partially composited color and opacity. The volume is then represented using the set of segments, and rendering this representation involves casting rays and intersecting the segments, which is faster than casting rays through the original volume data. This project aims to propose a novel technique to render VDIs in stereoscopic virtual reality. The primary challenges are two-fold: 1. A method needs to be devised for the stereoscopic rendering of VDI representation which is view-dependent and generated for a single camera viewport, and 2. An eye-tracking-based foveated rendering technique needs to be applied to render the VDIs at sufficient frame rate. This project will require programming GLSL shaders and programming in Kotlin or C++.

    • Sandbox App Development Tianfang Lin

      In a previous student project a sandbox augmented with a kinect and a beamer was built and a software was developed that allows interactive exploration of terrain flooding where the sand is used as terrain input and the beamer to show a flooding simulation on top of the sand surface. 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 Looking Glass, Stefan Gumhold

      The successful histocaching app developed in a previous student lab provides a playful access to historic content in the current context by overlaying historic photographs in an augmented reality mode of a cellphone app. The original app provides historic content on the main campus of TU Dresden. In order to bridge distances to other Campus parts or other historic cites in Dresden or anywhere on the earth, this topic shall examine an augmented reality looking glass. Distant content should be made accessible based on the relative direction and distance to the user. The topic includes setting up example content, developing a distant content navigation as well as selection concept as well as an evaluation of how well the spatial context is understood by the user.

    • Immersive Storytelling for point cloud based scenes,  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, 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.

    • AR treatment of anxiety disorders, Stefan Gumhold
      Current AR technology has reached a state where useful applications are possible. One of the most important applications of AR is the treatment of axiety disorders, which include claustrophobia, fear of height or speaking in public or spiders, etc. In this student work for one disorder a AR training tool should be developed and evaluated.

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

  • Rendering

    • Volume rendering the Adaptive Particle Representation (APR) Aryaman Gupta

      The Adaptive Particle Representation (APR) is a content-adaptive representation of 3D volume data. By adapting the resolution of the representation to the contents of the data, the APR has been shown to yield orders of magnitudes of benefit in the storing and processing of sparse volume data. This has led to its increased usage, particularly towards storing and processing 3D microscopy images. However, there currently exists no method and no tool for volume rendering the APR. This project explores latest work in volume rendering Adaptive Mesh Refinement data structures and applies them to the rendering of the APR. Potential techniques to explore include the use of Ray-Tracing cores to accelerate raycasting. The final goal of the project is to produce a tool that can be used by existing users of the APR to visualize their data. The project will require programming in GLSL and/or Vulkan or Optix Ray Tracing APIs.

    • Rendering Based Normal Estimation for Pointclouds, Tianfang Lin
      Normal estimation on pointclouds typically is based on a neighborgraph construction, plane fitting and global normal orientation. In this thesis an alternative approach should be investigated that renders the pointcloud to the depth buffer, estimates normals oriented towards the viewer from the depth buffer and then transfers the normals to the points in the pointcloud. Optionally, this approach can be integrated into an immersive VR-based normal estimation tool.

    • Accelerating Virtual Human Rendering with Tensor Cores, 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.

  • Virtual Humans
    • Remote Avatars
      Capturing, Transmission and Rendering of Human Avatars for Remote Collaboration, Julien Fischer
    • [RESERVED] Analyzis and Visualization of the AMASS Pose and Shape Space, Julien Fischer
      The AMASS dataset unifies a wide variety of human pose and shape datasets into a common representation in the form of parameters for the SMPL+H or SMPL-X parameterized human body models. In this thesis, the different shape and pose spaces which are present in the dataset should be analyzed in a first step. Afterwards, the results should be visualized, e.g. the variability of joint rotations based on age or sex.
    • Adaption of Human Animations According to Changes in Shape (and / or Environment), Julien Fischer
      The AMASS dataset contains a large number of human animations that were acquired through motion capture. In each frame, AMASS represents the human body as a set of SMPL+H or SMPL-X parameters. Since these parameters factor shape and pose, the shape of the animated body can be altered while the motion stays the same. However, changing the shape may lead to self-interpenetrations of the resulting body mesh. The goal of this thesis is to examine how, given a modified shape, the pose has to be adapted in order to prevent self-interpenetrations. 
    • Automatic creation of (full-body) identikits using expressive parameterized human body models, Julien Fischer
      • NOTE: This topic is suited exclusively for the module "Analysis of a research topic". Upon successful completion and assessed feasibility, this topic may be used for a thesis
      • Identikits (or phantom images) are currently created either by an artist or with the help of computer programs and templates. Parameterized human body models like SMPL-X may offer another alternative to create identikits through modification of their generative parameters. The goal of this topic is to analyze whether this approach is realizable and if yes, how. More detailed information on this topic will be made available soon.
    • Random Generation of Plausible Human Poses Julien Fischer
      Machine learning models require a lot of data during training. For use-cases regarding digital human twins, the AMASS dataset is often times used. While AMASS contains many human motion sequences, it's corpus of motion data is still limited. To generate even more possible training data, this topic should investigate how random human poses can be generated in a way such that the resulting poses are plausible (e.g. no self interpenetrations). Interesting sub-problems could also be how to restrict the randomly generated poses to certain types of movements (e.g. walking, rowing with the arms)
    • Physics-Based Simulations (PBS) for Clothed Virtual Humans, Kristijan Bartol
      Physics-based simulations are a well-known technique in computer graphics when it comes to particles and materials which are non-rigid. Human clothing is a good example of an application where PBS comes in handy. The clothing can be made of various physical materials, have various geometries, and can behave differently based on different body shapes. Given initial conditions (human body pose, shape, and clothing geometry on top of the body) and using PBS engines which solve differential equations of motion between the frames, we can simulate how clothing deforms over time based on given body motion. The task is to explore and get familiar with a PBS engine of choice for the problem of clothing simulation.
  • Scientific Visualization 

    • Guidance for Oblique Slicing, Stefan Gumhold

      Oblique slicing is a fast method for volume inspection and can be easily implemented in real-time. Still it is not used frequently in practice probably as it is not easy to adjust slices in an intuitive way. In this topic a guidance scheme shall be developed that makes slice adjustment for oblique slicing more intuitive. A possible approach is to analyse the sliced volume data in the vicinity of the current slice and to extract curve features that show how the data extends beyond the current slice. These curve features should be extracted on different scale levels showing very local to more global data behavior. Furthermore, slice adjustment shall be simplified by selection of a specific curve feature along which a 1D adjustment is possible.

    • Visualization of 3D Trajectories with Orientation, Stefan Gumhold
      Trajectories over 3D domains frequently arise in VR application tracking, air traffic and drone based logistic applications. Spatial mapping to line primitives ignores the orientation of moving objects. For planes and birds the orientation mostly corresponds to the direction of motion. But for VR player heads, drones or particles/objects in dynamic processes the orientation can change independent of the path. In this topic visual mappings shall be studied that show object orientation along 3D trajectories. Different mappings of orientation to visual attributes like color, texture orientation or height should be investigated and evaluated with respect to their effectiveness and expressiveness.

    • Visualization of 3D Trajectories of Agents that see, Stefan Gumhold

      Environment mapping, collision avoidance and navigation in unknown terrain demands for camera and/or 3D sensors on remote-controled or autonomous agents like cars, robots, or drones. In the analysis of the resulting evironment maps or in failure cases like crashes it is important to have access to the information what parts of the environment have been visible to the sensores of the agents for each time point. This motivates the visualization problem of trajectories with moving sensor frusti that shall be studied in this topic.

    • [RESERVED] Data Reduction for Trajectories Represented by Hermite Splines, Benjamin Russig
      Line data typically comes in densely sampled form, e.g. streamlines extracted from flow fields with small integration steps or densely sampled IMU-derived tracks of autonomous drones. When representing trajectories as Hermite Splines, their property of producing the minimal-tension curves between subsequent positions with associated tangent vectors promises a natural way of steering the amount of samples one can omit before storing a new node, assuming smooth motion and knowledge about the physical limits of the moving object with respect to acceleration. The goal of this project or thesis is to derive heuristics based on this property that can be used to simplify densely sampled trajectories in a greedy fashion (which is of special interest e.g. in the context of streaming), and compare the performance of such a method in terms of speed and approximation error to standard curve fitting approaches.

    • Interactive visualization of spiraling optical coherent tomography imagery, Stefan Gumhold

  • Particle tracking velocimetry in foams using radiography, Sascha Heitkam (pdf)

Topics already taken

  • Brain Twisters
  • Immersive Technology
  • Virtual Humans
    • Avatar-Configurator: Assigning clothes and assets to a virtual human avatar slides
    • Estimating Textures from an Image of a Clothed Person, Kristijan Bartol
      Estimating texture from a clothed person from an image is an advanced computer vision/graphics task of recovering UV maps between the texture images and the 3D mesh of the person. The 3D mesh of the person (geometry) is assumed to be known. On top of the human body, there are one or more layers of cloth, each with its own geometry and texture. The task is to estimate the texture maps for the visible parts of the image. For example, if the person is facing the camera, it is sufficient to estimate the texture of the visible, frontal parts. It is expected that the final solution includes a deep-learning approach, such as generative adversarial networks (GANs), but the specific architecture will be determined during initial discussions. Advanced differentiable computer graphics approaches such as inverse rendering could also be investigated.
  • Rendering

    • Remote Interaction for Visualization and Labeling

  • Scientific Visualization
    • Multi-channel Transfer Function Design for analyzing sliced microscopy data (within PoL), Stefan Gumhold
    • Immersive visualization of Morpheus' cellular automata simulation
      (see Morpheus project page),  Stefan Gumhold
    • Texture space grids with ticks and labels, Stefan Gumhold
    • Interactive Visualization of Extreme Subset Sets, Stefan Gumhold
      The automatic control of agents such as delivery drones or autonomous cars can be done with action policies that propose in each state the next action such as "turn left!", "break!", "accelerate!", ... . Action policies can be implemented with planing approaches or reinforcement learning. Here we consider action policies for a multi-goal objective such as delivery with drones to a multiple recipents within the negotiated delivery time. To analize multi-goal policies, one can extract all goal subsets of minimal size that cannot be fulfilled together. Alternatively, one can extract goal subsets of maximal size that can be fulfilled together. The goal of this student research topic is to develop an interactive tool for the inspection of extrem (minimal or maximal) goal subsets that scales to large goal sets. A good starting point for literature research with lots of pictures with potential subset visualization techniques can be found here.
    • Visualization of multi-variate 2D scalar fields
      In medical dignostics often different imaging modalities are necessary for a proper diagnosis. In this thesis different scalar field visualization techniques like height-fields, mapping to color, extraction of iso-contours or using lighting parameters should be investigated. Evaluation should examine which combination of visualization techniques is efficient and which combination is to be avoided.

    • Visualization of 2D Trajectory with Orientation, Stefan Gumhold

      Trajectories over 2D domains frequently arise in traffic and logistic applications. Spatial mapping to line primitives ignores the orientation of moving objects. For vehicles the orientation mostly corresponds to the direction of motion. But for particles or objects in dynamic processes the orientation can change independent of the path. In this topic visual mappings shall be studied that show object orientation along 2D trajectories. Different mappings of orientation to visual attributes like color, texture orientation or height should be investigated and evaluated with respect to their effectiveness and expressiveness.

  •  Computer vision topics

    • Real-time masking of water surfaces in video stream, Sebastian Vogt

    • Deep Family Fake: Synthetic generation of family pictures from images of a person using deep neural networks, Nishant Kumar 

    • MetaFuse: Meta-parameter optimization for balanced medical image fusion video  Nishant Kumar

    • Art Project with Anton Ginzburg:  Sebastian Vogt
      • Crash Video Synthesis By Simulation And Domain Transfer video

    • Likelihood-based Outlier Detection for Image Classification, Nishant Kumar          The deployment of reliable AI models is crucial for autonomous driving and medical imaging applications. In our recent work, we developed a likelihood-based approach for detecting outlier box features in an object detection setup. This project aims to extend the work to an image classification problem on standard image benchmarks. The training scheme related to our outlier-aware object detection approach will be discussed before the start of the project. 

    • Multimodal Medical Image Fusion for clinical prognosis, Nishant Kumar                  Multimodal Brain imaging, such as DWI and T2-Flair, are two common MRI modalities that provide crucial insights into tumor presence. For example, the high-intensity signal on DWI indicates that the tumor has solid components, i.e., enhancing tumor. In contrast, a high-intensity signal on T2-Flair indicates that the tumor contains more liquid components and is already necrotic. As the tumor grows, it becomes necrotic and liquefied due to insufficient blood supply. Clinicians perform pre-operative imaging to judge the extent of the tumor in the brain before conducting the surgical intervention. Subsequently, they perform post-operative imaging to analyze any tumor residual and anticipate the likelihood of tumor recurrence. However, for clinicians, the visualization of contrasting tumor features in a single fused image for both pre and post-operative images should help better comprehend the possibility of tumor recurrence. Therefore, this project shall develop an efficient approach that provides meaningful fused images viable for clinical interpretation and post-operative decision-making. The volumetric MRI datasets and the tumor segmentation labels will be provided before the start of the project.

    • Physics-informed Invertible Neural Networks, Nishant Kumar                                The project aims to extend the data-driven Invertible Neural Networks (see recent work)  by incorporating information from the physical law behind the simulation, to potentially improve the reconstruction of the conductivity map from the magnetic flux density. The project shall formulate the mathematical equation representing the underlying physical law and specifying appropriate boundary conditions. Additionally, the loss function that usually comprises only the data fidelity term in INN shall be extended with the physics-informed term, thereby enforcing the physical law as a constraint and ensuring adherence to the underlying physics. Data and the baseline code will be provided before the project begins.

    • Quantization-aware Invertible Neural Networks, Nishant Kumar                              Our recent work employed Random Error diffusion to quantize the continuous-valued conductivity maps obtained from the pre-trained INN. However, this algorithm only yields an ensemble of binary conductivity maps, lacking a unique solution. To address this, the project shall explore integrating error diffusion principles into the training process using quantization-aware techniques. The implemented approach shall introduce quantization operations during training, ensuring the network becomes robust to quantization errors. By doing so, the approach should handle quantization in a principled and differentiable manner, making them more suitable for training. Data and the baseline code will be provided before the project begins. 

    • Convolutional Neural Networks-based Reconstruction of Conductivity Maps, Nishant Kumar                                                                                                                Our recent work used fully connected layers as the learnable function of the Invertible Neural Network (INN) to reconstruct conductivity maps from magnetic flux density in electrolysis cell simulations. Since both variables are image-based, and Convolutional Neural Networks (CNNs) are generally suitable for image-related tasks due to their ability to capture local patterns like edges and textures, the project aims to investigate whether incorporating CNNs within INN architecture provides better solutions than fully connected layers based INNs. Hence, the project shall explore suitable CNN + INN architectures to solve the task without overfitting. Data and the baseline code will be provided before the project begins.
    • Reconstruction of High-resolution Conductivity Maps using Invertible Neural Networks, Nishant Kumar                                                                                             Our recent work demonstrated that Invertible Neural Networks (INNs) can reconstruct conductivity maps from magnetic flux density data in electrolysis cell simulations. The INN was trained using low-dimensional magnetic flux density from a single spatial component to reconstruct higher-dimensional conductivity maps. This project will investigate if utilizing additional flux density components, given a fixed number of sensors, allows for reconstructing a higher-resolution conductivity map. To achieve this, a detailed examination of the INN architecture and its meta-parameters must be conducted to avoid overfitting. Data and the baseline code will be provided before the project commences. 

  • User Guided Image Restauration with Invertible Neural Networks, Stefan Gumhold
    Invertible Neural Networks (INN) can map data distributions invertably to a uniform Gaussian distribution in a latent space of the same dimension as the input space. In this work INNs shall be used for image restauration by generating for a given corrupted input image a trajectory in latent space based on gradient ascent of the Gaussian probability distribution. Mapping the trajectory back to image space yields an image morph from the destorted image into the image manifold of the distribution learned by the INN. A user interface shall be provided to allow for interactive exploration of this image morph that can be used for user guided image restauration.          

Zu dieser Seite

Stefan Gumhold
Letzte Änderung: 23.07.2024