Multi-view stereo via geometric expansion and depth refinement

Multi-view stereo, which aims at reconstructing 3D models from series of images, has always been one of the important subjects in the field of robot vision. In this paper, we propose a novel framework to recover a 3D point cloud from calibrated images. Firstly, we construct a sparse point cloud from images by feature matching and triangulation. Then, the sparse point cloud is geometrically expanded to a dense point cloud model by using a shape prior patch library. Finally, we employ a depth refinement procedure so as to recover the details of the surface. Hence, in this work, the most difficult task, dense matching construction across the images, can be avoided as much as possible. Experimental results demonstrate the effectiveness of our method. On the reconstructed 3D model, the elaborate details are recovered, and the noise can also be suppressed as well.