Generate Individually Optimized Blendshapes

Blendshape based animation is a technique commonly used to animate the face. To generate realistic animation for different faces, creating appropriate blendshapes for each different face is essential. There have been many attempts to find the production-level blend shapes. However, many existing methods mostly require a professional artist's intuition with manual intervention. In this paper, we present a novel approach to automatically generate individually optimized blendshapes from real-time captured facial expressions. The proposed method generates the blendshape from the captured face with two methods: linear regression and an autoencoder. Among results from two methods, we select the trained result that is more similar to the original face. The adopted blendshape could be used to animate the original face more naturally. In addition, the generated blendshape is utilized to retarget the original face animation to another face while preserving the original face's animation characteristic. Comparison of results by animating the face on the screen show linear regression is suitable for retargeting the facial expressions without using the complicated neural networks.