Deformation Capture and Modeling of Soft Objects

摘要

We present a data-driven method for deformation capture and modelingrnof general soft objects. We adopt an iterative framework thatrnconsists of one component for physics-based deformation trackingrnand another for spacetime optimization of deformation parameters.rnLow cost depth sensors are used for the deformation capture, andrnwe do not require any force-displacement measurements, thus makingrnthe data capture a cheap and convenient process. We augmentrna state-of-the-art probabilistic tracking method to robustly handlernnoise, occlusions, fast movements and large deformations. The spacetimernoptimization aims to match the simulated trajectories withrnthe tracked ones. The optimized deformation model is then used tornboost the accuracy of the tracking results, which can in turn improvernthe deformation parameter estimation itself in later iterations. Numericalrnexperiments demonstrate that the tracking and parameterrnoptimization components complement each other nicely.rnOur spacetime optimization of the deformation model includes notrnonly the material elasticity parameters and dynamic damping coefficients,rnbut also the reference shape which can differ significantlyrnfrom the static shape for soft objects. The resulting optimizationrnproblem is highly nonlinear in high dimensions, and challengingrnto solve with previous methods. We propose a novel splitting algorithmrnthat alternates between reference shape optimization andrndeformation parameter estimation, and thus enables tailoring thernoptimization of each subproblem more efficiently and robustly.rnOur system enables realistic motion reconstruction as well as synthesisrnof virtual soft objects in response to user stimulation. Validationrnexperiments show that our method not only is accurate, but alsorncompares favorably to existing techniques. We also showcase thernability of our system with high quality animations generated fromrnoptimized deformation parameters for a variety of soft objects, suchrnas live plants and fabricated models.