Deep rigging : automatic character skinning for animation

Abstract

Animation typically involves attaching a skeleton to a character mesh with attachment weights and moving the skeleton's joints over time. Determining these attachment weights can be arduous and time-consuming, so we present a neural network architecture for producing these weights. During training, our architecture produces images of an input character deformed to match a series of reference poses. This deformation is performed inside a differentiable rendering pipeline, which utilizes affine transformations and attachment weights generated by independent networks. At test time, only part of our architecture needs to be evaluated to generate attachment weights for a given input character image. Our architecture adapts well to non-synthetic datasets and settings, demonstrating its extensibility and versatility.