Physical parameter prediction by embedding human perceptual parameter for 3D garment modeling

Seongmin Lee, Woojae Kim, Sewoong Ahn, Jaekyung Kim, Sanghoon Lee

Abstract: To model garments into a virtual environment, it is crucial to predict the physical parameters of the simulated model. However, it is troublesome for a user or technical director to intuitively reflect their aesthetic intention using physical parameters. In this paper, we propose a framework that predicts various physical parameters (e.g., stretch resistance, bend resistance, ...) by embedding human perceptual parameters (e.g., wrinkly, stretchy,...) in multi-task learning (MTL) perspective. By predicting both physical and perceptual parameters, we can effectively solve this problem, and can give an important cue to model a 3D garment maximizing users visual presence. Furthermore, by taking a class activation mapping method, our model seeks the intermediate visual understanding of physical and perceptual parameters. Through the rigorous experiments, we demonstrate that the predicted physical and perceptual parameters agree with subjective values.