Abstract

This paper introduces connectivity preserving constraints into spatio-temporal multi-view reconstruction. We efficiently model con- nectivity constraints by precomputing a geodesic shortest path tree on the occupancy likelihood. Connectivity of the final occupancy labeling is ensured with a set of linear constraints on the labeling function. In order to generalize the connectivity constraints from ob jects with genus 0 to an arbitrary genus, we detect loops by analyzing the visual hull of the scene. A modification of the constraints ensures connectivity in the pres- ence of loops. The proposed efficient implementation adds little runtime and memory overhead to the reconstruction method. Several experiments show significant improvement over state-of-the-art methods and validate the practical use of this approach in scenes with fine structured details.