Bilinear Kernel Reduced Rank Regression for Facial Expression Synthesis
People
- Dong Huang
- Fernando de la Torre
Abstract
Facial Expression Synthesis (FES) has been a flourishing area of research in the last few years driven by applications in character animation, computer games, video conferencing and human computer interaction. In this paper we propose a photorealist facial expression synthesis method based on Bilinear Kernel Reduced Rank Regression (BKRRR) that learns a high dimensional appearance mapping between a neutral image and a variety of expressions (e.g. smile, surprise, squint). The paper proposes three main novelties: (1) The use of BKRRR for FES. BKRRR effectively decouples the identity and expression factors. Also, several algorithms for learning the parameters in BKRRR from training data are evaluated. (2) A method to preserve subtle person-specific facial characteristics (e.g. wrinkles, pimples). (3) An illumination normalization algorithm to effectively render new expressions that might differ from the illumination conditions on the training set. Experimental results on the pictures taken with a regular camera and CMU Multi-PIE database show the effectiveness of our approach.
Citation
|
Dong Huang and Fernando de la Torre
Bilinear Kernel Reduced Rank Regression for Facial Expression Synthesis The 11th European Conference on Computer Vision, 2010 [PDF] [Bibtex] |
Results
Acknowledgements and Funding
This research is supported by:
Copyright notice
| Human Sensing Lab |