2D-to-3D conversion of single-view video based on view-distance classification

Abstract

In this thesis, we propose a 2D-to-3D conversion system based on view-distance classification. An input image frame is transformed by using $\It{Fourier transform}$, and feature vectors are obtained by $\It{principle component analysis}$ (PCA). Based on these feature vectors, the image view is categorized into $\It{close-up view }$, $\It{long view with vanishing point}$, or $\It{ long view without vanishing point}$. According to the view classes, different methods are applied to depth map generation, which is needed for stereoscopic video generation. Experimental results show the final depth map generated by each technique, and indicate our system is useful for various 3D applications.