Learning an Intrinsic Garment Space for Interactive Authoring of Garment Animation

摘要

Authoring dynamic garment shapes for character animation on body motionis one of the fundamental steps in the CG industry. Established workflowsare either time and labor consuming (i.e., manual editing on dense frames with controllers), or lack keyframe-level control (i.e., physically-based simulation).Not surprisingly, garment authoring remains a bottleneck in manyproduction pipelines. Instead, we present a deep-learning-based approach forsemi-automatic authoring of garment animation, wherein the user providesthe desired garment shape in a selection of keyframes, while our systeminfers a latent representation for its motion-independent intrinsic parameters(e.g., gravity, cloth materials, etc.). Given new character motions, the latentrepresentation allows to automatically generate a plausible garment animationat interactive rates. Having factored out character motion, the learnedintrinsic garment space enables smooth transition between keyframes ona new motion sequence. Technically, we learn an intrinsic garment spacewith an motion-driven autoencoder network, where the encoder maps thegarment shapes to the intrinsic space under the condition of body motions,while the decoder acts as a differentiable simulator to generate garment shapes according to changes in character body motion and intrinsic parameters.We evaluate our approach qualitatively and quantitatively on commongarment types. Experiments demonstrate our system can significantly improvecurrent garment authoring workflows via an interactive user interface.Compared with the standard CG pipeline, our system significantly reducesthe ratio of required keyframes from 20% to 1 − 2%.