Synthesizing Realistic Expressions in 3D Face Data Sets

This paper presents a robust method for synthesizing realistic expressions on 3D human face surfaces captured from frontal photographs with neutral expression. The generated facial expressions could be used to improve the performance of existing face identification systems, or to enhance human recognition. Firstly, the 3D face surface map is recovered using an analysis-by-synthesis approach based on a statistical model for encoding face shape information. Then a statistical discriminant model approach is employed to synthesize new facial expressions. The synthetic expression data is created by dividing a training set into two classes, for example smiling and frowning, and then finding the most discriminant direction between the classes. The expressions are applied to a human face by moving the surface points along this most discriminant direction. The resulting 3D models can be rendered under a variety of pose and illumination conditions. The key advantage of the proposed method is that expressions of varying degrees can be easily generated without having detailed changes in the 3D expression database. Besides altering human facial expressions, SDM could also be used to generate facial aging to aid the process of identifying missing children or adults after a time lapse. Moreover, synthetic biometric data contains 3D topology information, which is useful in analysing geometric facial shape changes over different facial expressions.