Dense 3D reconstruction from images by normal aided matching

Abstract

3D models play an increased role in today’s computer applications. As a result, there is need for flexible and easy to use measuring devices, that produce 3D models of real world objects. 3D scene reconstruction is a quickly evolving field of computer vision that aims at creating 3D models from images of a scene. Although many problems of the reconstruction process has been solved, the use of photographs as information source involves some practical difficulties. Therefore accurate and dense 3D reconstruction remains a challenging task. We discuss dense matching of surfaces in the case when the images are taken from a wide baseline camera setup. Some recent studies use a region growing based dense matching framework, and improve accuracy through estimating the apparent distortion by local affine transformations. In this paper we present a way of using pre-calculated calibration data to improve precision. We demonstrate that the new method produces a more accurate model. Key words: image based reconstruction,3D scene reconstruction, stereo, wide-baseline, dense matching.