Combining photometric normals and multi-view stereo for 3D reconstruction

In this paper a novel approach for 3D reconstruction of mid-sized objects is proposed that combines the advantages of multi-view stereo and photometric normal estimation. Images of an inspected object are taken in parallel with multiple synchronized cameras while the object is placed in an illumination setup that produces time-varying spherical gradient illumination for normal estimation. In contrast to existing approaches, the normal information is not only used to refine the preliminary stereo reconstruction. Instead, it is shown that the normal information allows to significantly increase the window size for stereo matching, which strongly improves robustness compared to classical window matching in the image domain. As a consequence, smoothness constraints, which are typical for stereo reconstructions, are no longer required. Our proposed method is comparably simple to implement and has a computation time of a few minutes. The evaluation on real and synthetic data demonstrates that detailed reconstructions with high accuracy are obtained.