Synthesizing Novel Views from an Uncalibrated Image Sequence

• Main Authors: Yi-Ting Chuang, 莊詒珽
• Other Authors: Cheng-Yuan Tang
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/13636962215695903078

碩士 === 華梵大學 === 資訊管理學系碩士班 === 96 === In this thesis, novel images from novel viewpoints are synthesized using uncalibrated images. There are two main parts in this thesis: (1) obtaining more stereo correspondences, and (2) novel views synthesis. Refer to the first part in this thesis. Given sparse 3D points (or stereo correspondences), there are two methods proposed to generate new stereo correspondences: (i) generating new stereo correspondences from 3D points, and (ii) generating new stereo correspondences from 2D points. In generating new stereo correspondences from 3D points, sparse 3D points can be regarded as seeds. New 3D points near 3D seeds are selected, and then these new 3D points are projected into images for verifying whether these new 3D points exist in images. In this thesis, ZNCC (Zero-mean Normalized Cross Correlation) and color histogram are used for verification purpose. In generating new stereo correspondences from 2D points, 2D corresponding triangular meshes are generated from sparse stereo correspondences. New 2D points within each triangular mesh are generated using the epipolar line. The procedure for verifying whether a new 2D corresponding point pair exists in two 2D corresponding point pairs is : (a) given a 2D point in a 2D triangular mesh in an image, an epipolar line in the corresponding 2D triangular mesh in the other image is generated. This epipolar line may cross two points in the corresponding 2D triangular mesh in the other image; (b) the corresponding point will be found by using ZNCC within a search region specified by these two crossing points. Refer to the second part in this thesis. Since stereo correspondences have been established, new images from novel viewpoints can be synthesized by using TriD which is the 3D display software developed by ITRI. The meshes from all 3D points are generated by using the Delaunay Triangulation. The texture mapping using these meshes can be done in TriD. Finally, we can use TriD to see the results of novel view synthesis from different viewpoints.