2D-to-3D stereoscopic conversion

Abstract

With the development of 3DTV technologies, the demand for 3D contents is increasing. Thus, the technique that converts two dimensional (2D) images to three dimensional (3D) images has become a big demand since there exist a number of images and videos captured in a conventional way. In this thesis, with introducing the related works for the stereoscopic conversion, we propose two novel stereoscopic conversion methods: one for static images which contain low depth of field and the other for common 2D soccer videos. The stereoscopic conversion of 2D images is straightforward once a depth map is provided. For the low depth of field images, we generate the depth map by extracting foreground regions through spatial-frequency components analysis on the image. For 2D soccer videos, we devise methods to generate the depth map which are different to shot types classified by soccer video analysis. In the long shot case, the depth map is generated with the assumption that the lower area in a frame is closer to viewers than the upper one is. For non-long shot case, we estimate the foreground region through skin color detection. In experimental results for each section, we present some pictorial results generated by our proposed algorithm, and subjective evaluation comparing the input videos and corresponding stereoscopic videos indicates that the proposed algorithm can yield the sense of depth from a single view images.