Abstract
We present a novel method to generate accurate and realistic clothingdeformation from real data capture. Previous methods for realistic clothmodeling mainly rely on intensive computation of physics-based simulation (withnumerous heuristic parameters), while models reconstructed from visualobservations typically suffer from lack of geometric details. Here, we proposean original framework consisting of two modules that work jointly to representglobal shape deformation as well as surface details with high fidelity. Globalshape deformations are recovered from a subspace model learned from 3D data ofclothed people in motion, while high frequency details are added to normal mapscreated using a conditional Generative Adversarial Network whose architectureis designed to enforce realism and temporal consistency. This leads tounprecedented high-quality rendering of clothing deformation sequences, wherefine wrinkles from (real) high resolution observations can be recovered. Inaddition, as the model is learned independently from body shape and pose, theframework is suitable for applications that require retargeting (e.g., bodyanimation). Our experiments show original high quality results with a flexiblemodel. We claim an entirely data-driven approach to realistic cloth wrinklegeneration is possible.