3D cartoon face rigging from sparse examples

We present a data-driven method for automatically constructing cartoonized 3D blendshapes of a subject’s face. Given a pre-defined blendshape template of the real facial expressions and corresponding cartoonized blendshape template created by an artist, we represent the blendshapes of an identity in the real and cartoon face spaces with the deformations of the blendshape template in each space and learn a mapping between the deformations in the two spaces. To this end, our method decomposes the deformations in each space into two parts: an identity-independent part that is represented with the deformation gradient of the blendshape template, and an identity-dependent part that is modeled by a low-rank linear model. We regress the linear model for the real expressions from a 3D facial expression dataset. An algorithm is then introduced to regress the mapping between the linear models in the two spaces from a small set of real expressions and their cartoonized counterparts. At run time, given the blendshapes of a subject’s real face and her 3D cartoon neutral face, our method automatically constructs the cartoonized blendshapes of the subject with the help of the cartoonized blendshape template and the learned mapping. Our method is user-independent and only requires a small set of 3D cartoonized expressions modeled by the artist for cartoon face rigging. We evaluate our method by creating cartoonized 3D facial animations for variant identities in two different artistic styles. The rigging results demonstrate that our method successfully preserves both artistic styles and personalized expressions of different identities.