This paper presents an algorithm for capturing the motion of deformable surfaces, in particular textured cloth. In a calibrated multi-camera setup, the optical flow between consecutive video frames is determined and 3D scene flow is computed. We use a deformable surface model with constraints for vertex distances and curvature to increase the robustness of the optical flow measurements. Tracking errors in long video sequences are corrected by a silhouette matching procedure. We present results for synthetic cloth simulations and discuss how they can be extended to real-world footage.
Volker Scholz, Marcus A. Magnor