Mesh ensemble motion graphs

Abstract

We describe a technique for using space-time cuts to smoothly transition between stochastic mesh animation clips while subject to physical noninterpenetration constraints. These transitions are used to construct Mesh Ensemble Motion Graphs for interactive datadriven animation of high-dimensional mesh animation datasets, such as those arising from expensive physical simulations of deformable objects blowing in the wind (see Figure 1). We formulate the transition computation as an integer programming problem, and use a novel randomized algorithm to compute transitions subject to noninterpenetration constraints. Asynchronous Mesh TransitionsCompressed mesh animations are attractive for real-time hardwareaccelerated playback of complex mesh deformation phenomena that is otherwise too expensive to compute on the fly [James and Twigg 2005]. Mesh ensembles with numerous distributed mesh elements, such as a garden of flowers or thousands of colliding flags driven by stochastic wind forces, provide a challenge for datadriven animation given the intrinsically high-dimensional dynamic phenomena. However, such examples are appealing candidates for interactive data-driven animation given the high cost of simulation using robust contact handling [Bridson et al. 2002]. Motion graph techniques [Kovar et al. 2002;Lee et al. 2002] provide the basic abstraction for data-driven animation, yet simply splicing together motion clips (to loop or transition within the dataset) leads to transition artifacts, since in high dimensional state spaces it is exceedingly rare to find close transitions, and synchronized transitions introduce undesirable errors with strong spatial correlation.To overcome these limitations, we allow asychonous transitions whereby each mesh subgroup can transition at a different frame so as to minimize its transition error (see Figure 2). In practice this is easily achieved by minimizing a separable objective function Φ( τ) that measures the smoothness of the transition as a function of each of the G mesh groups' transition offset times, τ ∈ Z G . Given a transition τ, we exploit tree-structured scene kinematics to reconstruct the scene shape during the asynchronous transition. Figure 2: Asynchronous transitions allow scene components to transition at slightly different times-yellow indicates mesh groups undergoing transitions-and thus avoid transition artifacts by (i) reducing total transition error, and (ii) diffusing group transition errors in both space and time. Noninterpenetration Constraint ProgrammingAlthough asynchronous mesh ensemble transitions are conceptually similar to space-time cut techniques used for video-based rendering [Kwatra et al. 2003], mesh animations have the added complexity of a 3D embedding for which not all shape configurations are feasible due to interpenetrations. Unfortunately, allowing mesh groups to transition any time they desire can lead to low-quality animations with substantial and unsightly interpenetration artifacts. We address this by minimizing the transition cost ...