Depth map generation of single-view images based on view-distance classification

Abstract

In this thesis, we propose a method of depth map generation based on view-distance classification. An input image is transformed by using Fourier Transform, and the image is categorized into close-up view, long view with vanishing point, and long view without vanishing point using 2-step Support Vector Machines (SVM) whose feature vectors are the spectra of the input image. In the close-up view case, the depth map is generated using an iterative reversible graph cut method based on saliency maps. In case of long view with vanishing point, the depth map is generated by a vanishing point detection technique. Finally, in the long view without vanishing point case, we use a sky detection algorithm and a gradient plane for depth map generation. We estimate the accuracy of the proposed image classification algorithm through experiments. Subjective evaluation is also conducted to assess our proposed system about depth map generation. Experimental results indicate that our system is competitive.