Towards Accurately Extracting Facial Expression Parameters

Existing methods extract facial expression parameters from captured expressions data in two steps. First, head absolute orientation is acquired, the captured expression is then transferred from world coordinates to local coordinates. Second, expression parameters are extracted from local coordinate data. However, above methods results to severe error accumulation. In the second step, the error of head absolute orientation is amplified when extracting expression parameters, which results to error accumulation of the synthesized facial expression. In this paper, we propose an optimization-based parameters extraction method to prevent large error accumulation. In our optimization model, we use the error generated from synthesized and captured expressions as optimization objective. The head absolute orientation and expression parameters were regarded as optimization variables, which were simultaneously computed through optimization. The optimization scheme reduces error accumulation effectively. The experiments have shown that, in the case of linear blend-shape expression parameterization, the proposed method was able to promote the accuracy and efficiency relative to existing methods; in the case of nonlinear blendshape parameterization, our method was able to synthesize even higher accuracy results than the previous work. Keywords— computer facial animation, expression capture, blendshape weights extraction, optimization, error accumulation