A novel 2D to 3D conversion method using 3D affine geometry and color consistency

Abstract

Recently, a large number of 3D content has been introduced and a variety of 3D products have been commercialized to the market. However, the number of valuable 3D content is still insufficient to attract public attention. In order to solve the limited supply of 3D content, various 2D to 3D conversion techniques have been studied by many research groups and corporations. Because most of the commercialized 3D display systems represent 3D images by simultaneously showing multi-view images, it is an important process to generate multi-view images from a single-view sequence with high-quality depth information in the 2D to 3D conversion. In this thesis, a novel method for 2D to 3D conversion is proposed, which largely consists of two parts: 3D-affine-geometry-based sparse depth computation and sparse to dense depth map generation using triangulation and color consistency. The proposed 2D to 3D conversion method can be applied to the sequences having static scenes with moving cameras since correspondences among several frames are required to obtain 3D affine geometry information. Basic concept of the proposed sparse depth computation step is that if the required virtual images could be limited to horizontal parallax images for the original images, it is possible to compute theoretically true disparity values by using only 3D affine geometry information, which has the same meaning as infinite homography or plane at infinity. In the 3D-affine-geometry-based sparse depth computation step, first, positions of the corresponding feature points in each horizontal parallax image plane are estimated using computed infinite homography. After that, infinite homography is updated to be applied to the subsequent frames and to maintain consistent baseline distances between the original and the generated horizontal parallax images through the sequence. The main advantage of the proposed 3D-affine-geometry-based sparse depth computation over previous 3D metric-structure-from-motion...