Preview

PDF, English Download (74MB) | Terms of use

Abstract

Scene reconstruction from camera images is a challenging task which benefits many applications, from visual effects to virtual and augmented reality applications (VR/AR) over industrial quality inspection to robotics and autonomous driving. This thesis is composed of four contributions to the field of passive 3D scene reconstruction: a) A fractal, passive, self-identifiying calibration target which provides a high number of calibration points independent of magnification. b) A general, ray-based calibration approach which allows highly accurate calibration of central and non-central single and multi-camera setups, from a passive calibration target. The approach is capable of estimating the target imperfections which enables fabrication of the calibration target via simple printing, where previously an active target had to be employed. c) A light field depth estimation method, which exploits the rich constraints from the many views of a light field capture to improve depth estimation. To this end an improved occlusion model is introduced and the resultant error terms are locally optimized using a fast patch-match based optimization scheme. The accuracy of this approach improves upon the previous state-of-the-art, as demonstrated in a multi-metric light field evaluation benchmark. d) Finally, the light field depth estimation approach is extended to exploit polarization cues captured with a light field polarization camera, which improves the reconstruction quality on smooth, glossy surfaces.