Facial Retargeting by Adding Supplemental Blendshapes

Abstract

This paper introduces a novel method to add a minimal set of missing blendshapes to automatically improve the quality of the retargeting result. Our approach compares the expression spaces defined by the source and the target to determine which frames need to be corrected. The expression spaces are represented on the principal axes extracted from the target blendshapes. Those frames that cannot be sufficiently reconstructed by the current set of blendshapes will be augmented by a set of automatically generated blendshapes. The new blendshapes are determined to minimize error between the source animation and the retargeting result. We performed experiments to compare the results created by a basic set of blendshapes and by our method. Our method effectively reduces error between the source and target animation, and produces much visually improved target animation without relying on manual intervention.