3D Reconstruction From Multiview Based On SIFT

We only get 2D images by capturing 3D objects in the world, but we can learn 3D information of objects from images. Getting 3D models of objects is one of the most important fields of Computer Vision. We can analyze and process corresponding information of an object by reconstructing its 3D model, which has been widely applied in medical imaging, robot navigation, virtual reality, topographical reconnaissance.Feature points detection and matching are vital for 3D reconstruction. We have made compare among Scale Invariant Feature Transform (SIFT), Harris, Smallest Unvalued Segment Assimilating Nucleus (SUSAN), and analyzed their capabilities of conserving spatial structures after matching. Many experiments have proved that SIFT and its matching algorithm are prior to Harris and SUSAN as well as their matching algorithms. Besides, we have improved matching algorithm of SIFT based on the fact that matched points are under the constraint of Epipolar Geometry, which can improve the accuracy of feature matching and make the matching algorithm more effective.This paper presents a multiview reconstruction algorithm based on calibrated cameras. The algorithm is based on the two-view reconstruction, so we will get multi-group 3D points. The fact that every 3D point relates to the only 2D point from a different view can help us figure out the correspondences among different groups of 3D points, and then multiview 3D reconstruction can be achieved.The new algorithm of 3D reconstruction based on multiview is easy to implement. Several experiments results prove that original appearance of object can be regained better.