Vorschau

PDF, Englisch Download (25MB) | Nutzungsbedingungen

Abstract

In traditional light-field analysis, images have matched spectral content which leads to constant intensity on epipolar plane image (EPI) manifolds. This kind of light field is termed homogeneous light field}. Heterogeneous light fields differ in that contributing images may have varying properties such as exposure selected or color filter applied. To be able to process heterogeneous light fields it is necessary to develop a computation method able to estimate orientations in heterogeneous EPI respectively. One alternative method to estimate orientation is the singular value decomposition. This analysis has resulted in new concepts for improving the structure tensor approach and yielded increased accuracy and greater applicability through exploitation of heterogeneous light fields.While the current structure tensor only estimates orientation with constant pixel intensity along the direction of orientation, the newly designed structure tensor is able to estimate orientations under changing intensity. Additionally, this improved structure tensor makes it possible to process acquired light fields with a higher reliability due to robustness against illumination changes. In order to use this improved structure tensor approach, it is important to design the light-field camera setup that the target scene covers the ±45° orientation range perfectly. This requirement leads directly to a relationship between camera setup for light-field capture and the frustum-shaped volume of interest. We show that higher-precision depth maps are achievable, which has a positive impact on the reliability of subsequent processing methods, especially for sRGB color reconstruction in color-filtered light fields. Aside this, a global shifting process is designed to overcome the basic range limitation of ±45° to estimate larger distances and to increase additionally the achievable precision in light-field processing. That enables the possibility to research spherical light fields, since the orientation range of spherical light fields typically overcomes the ±45° limit. Research in spherically acquired light fields has been conducted in collaboration with the German Center for Artificial Intelligence (DFKI) in Kaiserslautern.