Depth Estimation from Single Image for 2D-to-3D Conversion

Abstract

Images compress three-dimensional visual data to two-dimensional which exclude the third dimension of depth. Many different approaches to estimate depth information from single images have been taken towards solving this problem and significant progress has been made on the field during the last decade. But, the intrinsic limitation of these approaches is the loss of depth characteristics during the projection of the scene into the image plane. An advantage of these approaches is the relatively low amount of operations needed to process one single image, instead of two or more. However, even though many depth map generation algorithms have been developed, accurate estimation of depth information from natural scenes still remains problematic. To confront this issue, an approach is proposed in this dissertation to estimate accurate depth maps by using information from a single small-scale blurred image. The method takes the advantage of sharpness map to estimate the amount of spatially varying subtle defocus blur. By using edge information and applying a joint bilateral filter, a sparse depth map is obtained. Then, matting Laplacian interpolation algorithm is applied to attain a full depth map. This dissertation also introduces a new technique for converting the two-dimensional image back to three-dimensional with the help of recovered depth map. The evaluated depth estimates are added back to the corresponding two-dimensional (2D) image to give a three-dimensional (3D) impression to it. The robustness of proposed depth map estimation techniques are tested by several experiments and results are compared with the other well-documented methods, visually and quantitatively. When compared with J. Shi et al. [44], the average decrease in relative error, log10 error and root mean square error for real images used in this dissertation is 0.472, 0.1975, and 0.1952 respectively. The obtained results validate the effectiveness of proposed depth map estimation approach.