An interesting aspect of AAMs is their ability to easily interpolate the model that we trained our images on. We can get used to their amazing representational power through the adjustment of a couple of shape or model parameters. As we vary shape parameters, the destination of our warp changes according to the trained shape data. On the other hand, while appearance parameters are modified, the texture on the base shape is modified. Our warp transforms will take every triangle from the base shape to the modified destination shape so we can synthesize a closed mouth on top of an open mouth, as shown in the following screenshot:
This preceding screenshot shows a synthesized closed mouth obtained through active appearance model instantiation on top of another image. It shows how one could combine a smiling mouth with an admired face, extrapolating the trained images.
The preceding screenshot was obtained by changing only three parameters...