GarNet : A Two-stream Network for Fast and Accurate 3 D Cloth Draping Supplementary Material

Below, we provide further details about our dataset. Physics Based Simulation (PBS): Using the same notation as in the paper, our PBS ground truth was obtained as follows: Given the template garment mesh M, and the body mesh B, we obtain the ground-truth draping result G of M onto B via physics simulation. The physics engine is based on NvCloth [4] and incorporates a number of physical properties and forces acting on the garment, such as stretching, bending, gravity and dragging, while preventing garment-body interpenetration. As M is usually modeled in the A-pose, shown in Fig. 1, while B can have an arbitrary pose, we first warp B to the same pose as M. This pre-processing step is necessary for a good initialization of the physics engine. The warping process is achieved via Dual Quaternion Skinning (DQS) [3]. We then automatically compute the skeletons and blending weights for each body mesh B and template garment mesh M using the Pinocchio Auto Rig framework [2]. After warping B to match the pose of M, we gradually un-warp B to the original body pose. The garment is deformed accordingly during this process. Depending on the difference between the poses, this process can take from 100 to 200 iterations. To speed up the simulation, we use the GPU solver of NvCloth, which can only handle 500 triangles. Therefore, we first simulate the garment on a decimated 500-facet version of the body mesh on the GPU and then improve the draping results at full mesh resolution on the CPU. Dataset Splits: Table 1 shows the number of bodies Figure 1: Template garmentM in the A-pose.