AbstractsComputer Science

The 3D reconstruction of a scene from 2D images is an important topic in the eld of Computer Vision due to the high demand in various applications such as gaming, animations, face recognition, parts inspections, etc. The accuracy of a 3D reconstruction is highly dependent on the accuracy of the correspondence matching between the images. For the purpose of high accuracy of 3D reconstruction system using just two images of the scene, it is important to nd accurate correspondence between the image pairs. In this thesis, we implement an accurate 3D reconstruction system from two images of the scene at dierent orientation using a normal digital camera. We use epipolar geometry to improvise the performance of the initial coarse correspondence matches between the images. Finally we calculate the reprojection error of the 3D reconstruction system before and after rening the correspondence matches using the epipolar geometry and compare the performance between them. Even though many feature-based correspondence matching techniques provide robust matching required for 3D reconstruction, it gives only coarse correspondence matching between the images. This is not sucient to reconstruct the detailed 3D structure of the objects. Therefore we use our improvised image matching to calculate the camera parameters and implement dense image matching using thin-plate spline interpolation, which interpolates the surface based on the initial control points obtained from coarse correspondence matches. Since the thin-plate spline interpolates highly dense points from a very few control points, the correspondence mapping between the images are not accurate. We propose a new method to improve the performance of the dense image matching using epipolar geometry and intensity based thin-plate spline interpolation. We apply the proposed method for 3D reconstruction using two images. Finally, we develop systematic evaluation for our dense 3D reconstruction system and discuss the results.