Light Field Analysis
In contrast to a traditional image, the 4D light field of a scene captures the
appearance from several densely sampled view points simultaneously. Due to recent
advances in camera technology, recording of even light field video has become
comparatively cheap, which has sparked enormous interest to leverage this kind
of data in computer vision applications.
We study the rich structure of the light field with the aim of accurately reconstructing 3D geometry, material properties, and lighting information. In particular with for scenes with complex materials, light field analysis allows to solve problems which are next to impossible to deal with using only few sparse images.
In addition to our research in light field analysis, we also investigate more traditional 3D reconstruction algorithms based on multiple images, video or depth camera streams. In particular, our interest lies in efficient methods for large scenes based on octree representations, which we extend by variational methods for fusion, including super-resolution of texture and geometry.
In collaboration with the department of biology and MPI for collective behaviour, we pursue algorithms for efficient multi-target tracking in particular for animal collectives. The aim of this research is to investigate and understand swarm behaviour, in particular how complex global behaviour emerges from only local observations. For this, it is necessary to know about precise motion and ideally body pose of each individual, to be able to reconstruct the signals which are the basis for decisions.
Variational Inverse Problems
In our studies, we typically aim at physics-based continuous image formation models, which lead to variational inverse problems. We investigate novel approaches to formulate regularizers and data terms, in particular for our complex light field data structures, and study modern convex optimization frameworks to design efficient solvers. Recently, we have focused in particular on implementations on irregular and multi-resolution grids.