Our research focuses on the development of Computer Vision (CV) and Machine Learning (ML) Systems. Amongst others we are working on the reconstruction of natural habitats, novel imaging techniques, algorithms to quantify the behaviour of animals and the integration of CV and ML into 3D printing and AR & VR applications. Some research projects are listed below.
Image-based tracking of animals in their natural habitats can provide rich behavioural data, but is very challenging due to complex and dynamic background and target appearances. We present an effective method to recover the positions of terrestrial animals in cluttered environments from video sequences.More information can be found here...
Next to various applications, we also work on foundational developments in deep learning algorithms and architectures. While this includes theoretical models of deep learning, our focus is on empirical investigations into interpretability, optimization, and architecture design. More information can be found here...
We develop advanced computer vision methods to support wildlife conservation using aerial imagery. Our focus lies in the detailed 3D reconstruction of natural habitats and the accurate estimation of visual animal biometrics using drone technology. These techniques are aimed at supporting ecologists and researchers to extract rich ecological information from drone-captured data, improving our understanding of wildlife and their environments. More information can be found here...
Optical Neural Networks (ONNs) are a novel compute paradigm for developing hardware accelerators for deep learning. ONNs have physical limitations that require deep learning architecture adaptations as well as novel training schemes, which we investigate in this project. More information can be found here...
The goal of this project is to develop closed-loop algorithms which make 3D printing as easy as printing a text document. In particular, we want to integrate a variety of sensors like cameras and microphones into these gantry robots and develop new machine learning and computer vision algorithms to monitor the printing process. More information can be found here...
Augmented and virtual reality have heavily influenced our ways to perceive and interact with data. In this project we investigate the usage of AR and VR technology within a clinical and medical context. In particular, we are developing novel computer vision, machine learning and computer graphics algorithms and environments for medical education. More information can be found here...
The analysis of large biomedical image datasets remains a significant challenge. Whether dealing with high-resolution whole slide images (WSIs) of tissue samples or extensive volumetric data, processing and extracting meaningful features are still computationally expensive and time-consuming tasks. More information can be found here...
The annotation of image data is often a time-consuming and expert-intensive task, particularly when working with new domains in the medical field. As a step toward making this process more efficient, we explore the use of eye-tracking technology as an additional input for neural networks. More information can be found here...
Other research projects (partially active, partially completed) can be found below.
We have developed a new FTIR-based Imaging Method (FIM), which is used to detect the contact surface between the arena and the animals to enable high-contrast and high-throughput behavioural analysis. This system in combination with the associated tracking software FIMTrack.
We have implemented Habitat3D, an open source software tool to generate photorealistic meshes from point clouds of natural outdoor scenes. Habitat3D offers a variety of different filtering, clustering, segmentation and meshing routines which can be assembled into pre-defined pipelines operating on either subsets (i.e. clusters) or complete clouds.
Our FTIR-based Imaging Method (FIM) results in an excellent foreground/background contrast so that internal organs and other structures are visible without any complicated imaging or labelling techniques. We demonstrate that FIM enables the precise quantification of locomotion features namely rolling behavior and muscle contractions and we demonstrate that FIM enables automatic in vivo heartbeat quantification of Drosophila melanogaster pupae.
We have developed techniques to segment and track of multiple cells in in vivo wounding. In particular, challenges like overlaps and clustered cell entities are addressed by introducing a statistical measure called hemocyte migration score.
If two cameras are employed to estimate the trajectories of identical appearing objects, calculating stereo and temporal correspondences leads to an NP-hard assignment problem. We study two different types of approaches: probabilistic approaches and global correspondence selection approaches.