H.-C. Lo and R. Chung (PRC)
Facial expression recognition, Motion in depth, Affine camera, View synthesis.
Visual recognition of facial expression is an important topic of affection computing that could be used to make machines appear more human-friendly. However, the recognition problem under an unknown motion in depth of the face is difficult because the viewpoints associated with the candidate image and the model (reference) image data could be very different, making it hard to match the model data directly with the candidate image. We explore a solution that, given a candidate image to recognize, first synthesizes from the model data (two or more model images) a novel view that is under a viewpoint more or less the same as that of the given image, and uses the novel view as a recognition template to compare with the given image. The solution assumes that the imaging process could be well approximated as an affine transformation. We present experimental results on real image data. We also demonstrate how the recognition solution could be used for driving the facial expression of an animation character automatically.
Important Links:
Go Back
