Abstract
To solve the disadvantages of the recent three-dimensional (3D) reconstruction methods, this paper presents a dense reconstruction algorithm which incorporates image features into the mesh subdivision. A preliminary model is deformed from statistical deformation model (SDM) to catch the roughest features, and later densified under a multi-resolution frame to achieve a dense model until all image features are used. In every round of this framework, the image features are detected, and then cut into planar points. At the same time, the transitional model is re-projected into views, and all images are segmented to triangles. So the planar points are constrained by these triangles, and the corresponding relations among these triangles are certain. Then some new spatial points are generated from these planar points and subdivide the transitional model. Later, the transitional model is re-arranged to get a right topology, and becomes denser. This method contributes three novelties: (1) there is an inherited relationship between the later points and the early points; (2) if the hierarchy of multi-resolution frame and the number of new spatial points at every round are set appropriately, the transitional model evolves with a correct topology; (3) the density of the final model is proportional to the information account of views. The experiment validates these virtues.
Acknowledgment
This research was jointly sponsored by Zhejiang Provincial Natural Science Foundation of China (Project No. Y1100075), Technological Project of Public Welfare Founded by Science Technology Department of Zhejiang Province (Project No. 2012C31G2250013), Shanghai Key Discipline Fund of Shanghai University (Project No. B 67) and Science Foundation of Wenzhou University, which is greatly appreciated by the authors.