Model-Based MVS point reconstruction of texture-less regions with epipolar constraints

ABSTRUCT: Point clouds are acquired with Structure from Motion/Multi-View Stereo (SfM/MVS) and laser scanning for Building Information Modeling (BIM). Although the SfM/MVS can generate dense point clouds, it is not easy to reconstruct texture-less regions because the SfM/MVS is based on feature-based image matching. Thus, in metal bridge measurements, point clouds are not generated in many texture-less regions such as the plane of the girder. Therefore, we propose a model-based MVS methodology with epipolar constraints using the intrinsic parameters and extrinsic parameters estimated with SfM processing. Our point cloud reconstruction approach consists of SfM, texture-less region selection with sparse point cloud back-projection, and dense point cloud generation with model-based MVS. We selected metal bridge girders as measured objects. Through our experiment, we confirmed that our methodology can reconstruct point clouds, even if measured regions are texture-less.