Multi-view3D Reconstruction Method

With the development of multimedia technology,3D reconstruction technology is widely used in the field of movie special effects, digital city,3D games. In recent years, multi-view stereo3D reconstruction is one of the most fundamental problems in computer vision.In this paper, first we introduce two commonly used camera calibration methods, and the Oct-tree based visual hull construction algorithm. Then we introduce two multi-view3D reconstruction methods. One is based on stereo matching in which the space is divided into voxels, another is based on photometric stereo. These two methods can work together to get a more precise model.The first method is based on the ray. The geometric representation is a discrete volume divided into voxels. A ray is generated for each pixel. NCC is used as the consistency measure of ray and voxel. When calculating NCC, the normal of the patch is sampled which makes NCC much more confident. Finally, a global optimization model based on factor graph is proposed. Thanks to the special property of the ray factor, we investigate an efficient algorithm based on the general loopy belief propagation, which has reduced the computational complexity from exponential to linear. Experiments show our approach is robust, the accuracy and completeness of the reconstructed models are good.The second method is based on photometric stereo. We use lighting information to calculate the normal, which is then to be used to optimize the single-view depth map or complete3D model. In this paper, we design a simple lighting pattern and introduce an algorithm to calibrate the ball. Then we introduce two algorithms to optimize the single-view depth map and complete3D model. Experiments show our approach doesn’t depend on the texture and can produce high quality details on the surface.