Dipl.-Inform. Lisa Brückbauer

Teaching

Background & Past Research:

In the first part of the PMDLumi project, we focused on developing new techniques for real-time rendering of light fields from 3D models. The standard approach of these techniques is based on a set of preacquired image-data. From these image data, new virtual views can be synthesized. Additional per pixel depth information, provided by a time-of-flight camera, is taken into account to improve image quality.

The light field data format as well as techniques for real-time acquisition and rendering of light fields have been developed using synthetic data. Synthetic data are currently extracted from a PMD-Simulator, which has also been developed by our group.

Our light field data format is based on a spherical approximation parameterizing camera space. For each sample position on the spherical arrangement a parabolic map of the opposite hemisphere, containing combined RGB and depth values per-pixel (RGBz), is stored. Both, the resolution of the camera parametrization as well as the image resolution of the parabolic map are adjustable. For camera space parametrization we provide hierarchical spherical arrangements containing either 12, 42, 162, 642 or even 2562 sample positions.

Our spherical light field renderer is based on a raycasting approach and runs on the GPU. It uses per-pixel depth information for efficient depth correction of rays. Image synthesis is then performed by rendering the smooth shaded faces of the polygonal camera sphere. For per-fragment interpolation an additional false-color is assigned to each vertex to establish per-pixel camera weights with individual polygons. Surface points are established very precisely for each fragment based on the parabolic maps stored with the nearby vertices.

Applying our raycasting approach, high-quality rendering results are achieved without any visible ghosting artifacts at high update rates of up to 53 frames per second. The rendering results show that simple objects can be reconstructed from as few as 12 light field samples. For more complex objects a sample count of 42 has proven to be sufficient in most cases.

For more details and results please refer to this page.

This research project is partially funded by grant KO-2960-6/1,2 from the German Research Foundation (DFG).

Current research:

Current research in the project is focused on relighting approaches. The main goal is to find new and efficient ways to allow for an interactive change of a light field's illumination in realtime. This is especially important in the context of object recognition. Being capable of relighting the existing image data in real-time will make the system insusceptable to varying lighting conditions.