Augmenting fast stereo with silhouette constraints for dynamic 3D capture

We propose and evaluate a method that combines a visual hull with multi-view stereo for the fast acquisition of dynamic 3D data. Both use the same set of cameras. The focus lies on high speed, as 3D has to be generated repeatedly, based on the output of multiple video cameras. The underlying principle is to bootstrap the system with quickly computed sparse stereo, constrained by silhouette information. We further show how this integrates with a fast plane sweeping algorithm, as our dense stereo algorithm of choice