AbstractsAstronomy & Space Science

This work is about the analysis of 4D light fields. In the context of this work a light field is a series of 2D digital images of a scene captured on a planar regular grid of camera positions. It is essential that the scene is captured over several camera positions having constant distances to each other. This results in a sampling of light rays emitted by a single scene point as a function of the camera position. In contrast to traditional images – measuring the light intensity in the spatial domain – this approach additionally captures directional information leading to the four dimensionality mentioned above. For image processing, light fields are a relatively new research area. In computer graphics, they were used to avoid the work-intensive modeling of 3D geometry by instead using view interpolation to achieve interactive 3D experiences without explicit geometry. The intention of this work is vice versa, namely using light fields to reconstruct geometry of a captured scene. The reason is that light fields provide much richer information content compared to existing approaches of 3D reconstruction. Due to the regular and dense sampling of the scene, aside from geometry, material properties are also imaged. Surfaces whose visual appearance change when changing the line of sight causes problems for known approaches of passive 3D reconstruction. Light fields instead sample this change in appearance and thus make analysis possible. This thesis covers different contributions. We propose a new approach to convert raw data from a light field camera (plenoptic camera 2.0) to a 4D representation without a pre-computation of pixel-wise depth. This special representation – also called the Lumigraph – enables an access to epipolar planes which are sub-spaces of the 4D data structure. An approach is proposed analyzing these epipolar plane images to achieve a robust depth estimation on Lambertian surfaces. Based on this, an extension is presented also handling reflective and transparent surfaces. As examples for the usefulness of this inherently available depth information we show improvements to well known techniques like super-resolution and object segmentation when extending them to light fields. Additionally a benchmark database was established over time during the research for this thesis. We will test the proposed approaches using this database and hope that it helps to drive future research in this field. Die vorliegende Arbeit besch¨aftigt sich mit der Analyse von 4D Lichtfeldern. Als Lichtfeld bezeichnen wir in diesem Zusammenhang eine Serie von digitalen 2D Bilder einer Szene die auf einem planaren regul¨aren Gitter von Kamerapositionen aufgenommen werden. Essenziell ist dabei die Aufnahme einer Szene mittels vieler Kamerapositionen konstanten Abstandes zueinander. Dadurch werden die von einem Punkt der Szene ausgehenden Lichtstrahlen als Funktion der Kameraposition abgetastet. Dadurch ergibt sich die bereits erw¨ahnte Vierdimensionalit¨at der Daten da, im Gegensatz zu einem …