Abstract

In this thesis we address the problem of automatically converting a video filmed with a single camera to stereoscopic content tailored for viewing using 3D TVs. We present two techniques: (a) a non-parametric approach which does not require extensive training and produces good results for simple rigid scenes and, (b) a deep learning approach able to handle dynamic changes in the scene. The proposed solutions both include two stages: depth generation and rendering. For the first stage, for the non-parametric approach we utilize an energy-based optimization, and for the deep learning approach a multi-scale convolutional neural network to address the complex problem of depth estimation from a single image. Depth maps are generated based on the input RGB images. We reformulate and simplify the process of generating the virtual camera’s depth map and present how this can be used to render an anaglyph image. Anaglyph stereo was used for demonstration only because of the easy and wide availability of red/cyan glasses however, this does not limit the applicability of the proposed technique to other stereo forms. Finally, we have extensively tested the proposed approaches and present the results.