To date, conventional color imaging has become highly sophisticated with images and videos of ultra-high resolution, high dynamic range and wide color gamut. Beyond these well-known features, we observe an introduction of computational methods into 2D imaging, enabled by the enormous performance of today's graphics processors. Thus, the next evolutionary step comprises computational methods utilizing 3D information within a capture volume.
However, there is still a severe lack of combined high-quality 2D and 3D capture technology in the consumer and even more in the professional domain. There are depth sensing technologies available, such as Time-of-Flight imagers. But they suffer from very low resolution, strong noise and the missing integration into a practical professional camera setup. Moreover, today's setups are limited because of complex 3D camera calibration processes, which are essential, when providing consistent RGBD data. Binocular RGBD camera approaches typically suffer from a fragile 3D calibration and mutual occlusion, which are difficult to resolve.
Given a set of RGBD data, an important element in today's research is the rendering of synthetic lens characteristics offering new flexibility. Thereby, real lens characteristics like defocused areas are eventually no longer irremovably incorporated in the recorded images.
State-of-the-art methods for lens synthesis and deblurring are limited as they require depth maps, which are not provided by professional cameras yet, and a tremendous amount of manual supervision.
Furthermore, the real-time availability of RGBD data allows for real-time interaction of computer-generated content and live action, as they can be spatially registered for the first time.
In this dissertation, a novel generalized monocular RGBD camera system is proposed. It integrates a high-quality RGB sensor and a depth sensor of different size, behind one main lens. Moreover, an in-depth evaluation of the main optical parameters, which imply the spatial consistency of such a camera approach, is given.
Using our camera, its proposed simplicity and the robustness of sensor registration is verified. The evaluations show that the viewpoints of both sensing systems are identical with respect to certain constraints. Moreover, solutions to acquired optical multi-spectral artifacts of our system are provided.
Furthermore, this work proposes essential sensor fusion methods for Time-of-Flight-specific noise reduction, effective depth map upscaling and global depth map denoising to finally arrive at a signal processing workflow delivering a highly enhanced amount of spatial and temporal consistency of the delivered RGBD data.
Based on this foundation, novel methods for synthetic cinematic depth-of-field synthesis as well as its inverse, the depth-guided deblurring process, are provided. Our approach deeply integrates knowledge about the camera system into the algorithms to overcome state-of-the-art methods in effectiveness and automatization.
Concluding, the applicability of our camera in a novel approach for seamless interaction of computer-generated content and live action, is demonstrated.
«
To date, conventional color imaging has become highly sophisticated with images and videos of ultra-high resolution, high dynamic range and wide color gamut. Beyond these well-known features, we observe an introduction of computational methods into 2D imaging, enabled by the enormous performance of today's graphics processors. Thus, the next evolutionary step comprises computational methods utilizing 3D information within a capture volume.
However, there is still a severe lack of combined h...
»