Reconstruction of Photo Consistent Shape from Multiple Images

• Main Authors: Tsai, Chang-Ming, 蔡昌明
• Other Authors: Tang, Yaw-Chung Gregory
• Format: Others
• Language: en_US
• Published: 2001
• Online Access: http://ndltd.ncl.edu.tw/handle/36763509159877290970

碩士 === 國立臺灣大學 === 資訊工程學研究所 === 89 === Reconstruction of 3D structures from multiple images has long been an important topic in computer vision. In the past, two of the most significant types of information being used for 3D reconstruction were the image-correspondence and the silhouette. In this thesis, we propose a new technique for 3D reconstruction, which exploits both of these types of information in an integrated manner. The 3D structure reconstructed in our work, which is called the minimal photo hull, is a type of photo-consistent shape, i.e., the shape reproduces the input images. In addition, the minimal photo hull satisfies the min/max optimization criterion. That is, it allows the maximal matching error associated with the color inconsistencies of the image correspondences to be the minimal. To apply the minimal photo hull for novel-view generation and 3D composition, we develop the view-dependent and view-independent texture mapping methods. On the one hand, the view-dependent novel views look more realistic; on the other hand, the view-independent texture mapping provides a consistent representation of a voxel-color structure. We further consider the compression of the voxel-color structure of a 360° object movie. A cylindrical structure is proposed as a preliminary study of the compact representation of voxel-color structures so that it can be compressed using existing 2D compression techniques.