Vorschau

PDF, Englisch Download (25MB) | Nutzungsbedingungen

Abstract

In this thesis we address supervised algorithms and semi-manual working steps which are used for scenarios where automatic computer vision approaches cannot achieve desired results. In the first part we present a semi-automatic method to acquire depth maps for 2D-3D film conversions. Companies that deal with film conversions often rely on fully-manual working steps to ensure maximum control. As an alternative we discuss an approach which uses computer vision methods to reduce processing time but still provides opportunities to interactively control the outcome. As result we receive detailed, smooth and dense depth maps with sharp edges at discontinuities. Part II, which presents the major contribution of this work, deals with human annotations used to assist ground truth acquisition for computer vision applications. To optimize this labour-intensive method, we analyse whether annotations created by different online crowds are an adequate alternative to running such projects with experts. For this purpose we propose different methods for improving acquired annotations. We show that appropriate annotation protocols run with laymen can achieve results comparable to those of experts. Since online crowds have much more users than typical expert groups used to run according projects, the presented approach is a viable alternative for large data acquisition projects.