Real-time rendering and acquisition of spherical light fields

摘要

Image-based rendering techniques have proven to be a powerfulalternative to traditional polygon-based computer graphics.This thesis presents a novel light field rendering techniquewhich performs per-pixel depth correction of rays forhigh-quality light field reconstruction. The technique storescombined RGB and depth values in a parabolic 2D texture forevery light field sample being acquired at discrete positionsin a uniform spherical setup. Image synthesis is implemented onthe graphics processing unit within a customized fragment program which extracts thecorrect image information from adjacent cameras for eachfragment by applying per-pixel depth correction of rays.This dissertation demonstrates that the presented image-basedrendering technique provides a significant improvement comparedto previous approaches. Two different rendering implementationsare explained which make use of the uniform parametrization tominimize disparity problems and ensure full six degrees offreedom for virtual view synthesis. While one renderingalgorithm implements an iterative refinement approach forrendering light fields with per-pixel depth correction, theother approach employs a raycaster which provides superiorrendering quality at moderate frame rates.Graphics processing unit based per-fragment depth correction of rays, used in bothimplementations, helps reducing ghosting artifacts to a nonnoticeable amount and provides a rendering technique thatperforms without exhaustive pre-processing for 3D objectreconstruction.The presented light field techniques open up for theimplementation of efficient and flexible rendering approaches.This work presents an efficient level of detail renderingapproach for light fields and introduces a flexible renderingtechnique for remote access to light field representation in aweb-based client-server application.For the acquisition of spherical light fields with per-pixeldepth a new acquisition system is presented which makes use ofrecent advances in 3D sensor technology to acquire combined RGBand depth images directly.