Completing sparse reconstruction in few strokes

We present a novel interactive framework for creating complete and accurate 3D models starting from low-quality results of multi-view stereo matching 3D reconstruction techniques. Our framework is motivated by the fact that even state-of-the-art solutions may provide very poor results, for example noise, missing parts, holes, in certain conditions. For example, litte overlap between images, bad lighting conditions, moving occluders, homoge- neous appearance, are some of the conditions that prevent those algorithms to obtain accurate and reliable reconstruction. In this paper we propose a frame- work that allows to take one such reconstruction and turn it into a complete model by requiring limited user interaction. The framework is based on two novel techniques that are the main contrbution of this work. Ths first is a multi-view segmentation algorithm that enables the user to select a region on a single image and propagate such selection jointly to the other images and the corresponding 3D points. The second is a GPU based algorithm for the reconstruction of smooth surfaces from multiple views that may incorporate user given hints on the type of surface. We show how the proposed framework may be resolutive in several situation where state-of-the-art method perform poorly or fail altogether.