Abstract
Multiple view reconstruction of human performance as a 3D video has advanced to the stage of capturing detailed non-rigid dynamic surface shape and appearance of the body, clothing and hair during motion [Aguiar et al. 2008; Starck and Hilton 2007]. Full 3D video scene capture holds the potential to create truly realistic synthetic animated content by reproducing the dynamics of shape and appearance currently missing from marker-based motion capture. However, the acquisition results are in an unstructured volumetric or mesh approximation of the surface shape at each frame without temporal correspondence, which makes the reuse of this kind of data more challenging than conventional mocap data. In this paper, we introduce a framework that automatically constructs motion graphs for 3D video sequences and synthesizes novel animations to best satisfy user specified constraints on movement, location and timing.