The growing availability of three-dimensional (3D) displays in the commercial sector and related trends in this field increase the demand for 3D content. Therefore, the development of tools for the generation of 3D content is of academic and commercial interest. This thesis proposes a semi-automatic 2D-to-3D depth propagation algorithm, which requires only little user-effort and delivers good quality results. To perform the propagation, the user only needs to assign depth by annotating scribbles on the first and last frame. The proposed algorithm builds upon a cost volume filtering-based video object segmentation algorithm and propagates the user-assigned depth values to all frames of the video. Our work focuses on several quality attributes including temporal consistency, edge-sharpness mismatches and depth changes over time. The proposed algorithm achieves improved temporal consistency by enhanced temporal filtering. Moreover, the algorithm accounts for the depth change of an object moving throughout the video by the incorporation of the depth order of the video. Evaluations show that the proposed algorithm requires minimal user interaction and generates spatio-temporally coherent depth maps with a perceptually consistent depth change for objects that change their depth in time. Moreover, a quantitative and visual comparison of the proposed algorithm to a related 2D-to-3D conversion algorithm is performed. It is shown that high-quality results can be achieved by using a robust optical flow estimation and a comfortable scribble annotation by the user.
T. Schausberger: "Temporally Coherent Cost Volume Filtering-based Depth Propagation in Videos"; Supervisor: M. Gelautz, N. Brosch; Institut für Softwaretechnik und interaktive Systeme, 2015; final examination: 04-14-2015.
Click into the text area and press Ctrl+A/Ctrl+C or ⌘+A/⌘+C to copy the BibTeX into your clipboard… or download the BibTeX.