Robust, Accurate and Weakly-Supported-Surfaces preserving Multi-View Reconstruction

Michal Jančošek
CMP Prague, Czech Republic

We propose a novel method for the multi-view reconstruction problem from a 3D point cloud. Surfaces which do not have direct support in the input 3D point cloud and hence need not be photo-consistent but represent real parts of the scene (e.g. low-textured walls, windows, cars, ground planes) are very important for achieving complete reconstructions. These difficult surfaces often coexist with highly textured easy surfaces which can be quite accurately reconstructed with standard techniques. If a difficult surface partially occludes an easy one, then the visible part of the easy surface can be seen as the background part of a silhouette image. This, similarly to Visual-Hull, can define the difficult surface i.e. weakly-supported surface. Our method trade in that fact and therefore is able to robustly reconstruct the weakly-supported-surfaces. We demonstrate an importance of these surfaces on several real world datasets. We compare our method to our implementation of the most relevant state-of-the art method and show that our method can considerably better reconstruct weakly-supported-surfaces while preserving thin structures and details on exactly the same level (and in the same computational time) as the competing method.