Facial Type, Expression, and Viseme Generation Josh McCoy, James Skorupski, and Jerry Yee
Introduction Virtual Human Faces –Hard to generate –Easy to criticize Motivation –Movies –Games Problems –Hand-made models take time –Physically-based models look weird
Contribution Data-driven facial face generation User-guided categorization Real-time pose generation from data
Related Work: Face Retargeting V. Blanz, C. Basso, and T. Vetter Reanimating Faces in Images and Video. –Use a morphable model to synthesize a 3D face of the 2D image. –Capture 35 scans of static face poses (expressions and visemes in neutral expression) from a source actor. –Find dense point-to-point correspondences –Retarget facial movements to the 3D face. –Render the 3D face back into the 2D image.
Related Work: Face Retargeting Problems –Does not generate new expressions that are not in the source data set. –Does not combine and retarget expressions and visemes together.
Related Work: Bilinear Model E. Chuang and C. Bregler Mood Swings: Expressive Speech Animation –Capture a video of an actor reading script under three different expressions (happy, angry, neutral) –Create a bilinear model, factoring expressions and visemes into two separate components. –Synthesize new facial movements with any expression and viseme.
Related Work: Bilinear Model Problems –Requires a full Cartesian product of facial expressions and visemes. –Does not generate new expressions that are not in the source data set. –Does not change the facial characteristics (identity). Pres Videos\Jerry\moodswings.mov Pres Videos\Jerry\moodswings.mov Pres Videos\Jerry\moodswings.mov
Related Work: Multilinear Model D. Vlasic, M. Brand, H. Pfister, & J. Popovic Face Transfer with Multilinear Models –Capture videos of 16 actors, each performing 5 visemes under 5 different expressions. –Create a multilinear model, factoring expressions, visemes, and identity into three separate components. –Synthesize new facial movements with any expression, viseme, and identity
Related Work: Multilinear Model Problems –Requires a full Cartesian product of facial expressions, visemes, and identity. –Limitations in the missing data imputation process. –Does not generate new expressions that are not in the source data set. Pres Videos\Jerry\vlasic-2005-ftm-sing.mp4 Pres Videos\Jerry\vlasic-2005-ftm-sing.mp4 Pres Videos\Jerry\vlasic-2005-ftm-sing.mp4
Methods Acquire and Categorize Learn Generate
Acquire and Categorize Three data sets are needed to fill the model space –Set of many neutral faces –Set of one face in many poses –Set of Visemes with reference face Vertex Correspondence User “rates” attributes of each face Video
Acquire and Categorize
Learn Expression deformation Viseme deformation Type deformation Reference Face Analyze each triangle and transform type separately
Learn Low-dimensional subspace (PCA) polygons individuals Compare each pose to reference face Principle Component Analysis (PCA) –Apply to each axis of variation –Analyze transformation of every face in mesh Infer variation of single attribute from combination of many
Generate Same sliders as categorization UI Generate any combination of attributes Runs in real-time
Results
Conclusion Realistic face poses from real-world basis data Arbitrary faces from sparse data set Future Work –Use high res data to drive low res morphing –Incorporate more biologically accurate face model