Synthesizing Segmented Mug-Shot Images for Recognition

For recognition, we select a single pose or $(t_{x},t_{y},t{z},\theta_{x},\theta_{y},\theta_{z},s_{y})$ value and use it exclusively to map the faces detected into a consistent view. The view we select is a frontal, mug-shot view which seems best for recognition purposes (although we could try synthesizing profile views, for instance and checking how recognition results vary with such data).

Since the faces will be projected into a standard, frontal view, using a pre-determined pose, a fixed cropping or segmentation of the 2D projection can be performed and reapplied to every new face undergoing the projection. Thus, we can automatically generate consistent frontal, segmented mug-shot views of any individual as long as we specify the four 2D anchor points corresponding to the eyes, the mouth and the nose. An example of a synthesized mug-shot face is shown in Figure .

  

Figure 4.16: A synthesize mug-shot image of U.S. President Ford.