I used the following images for the first set of experiments.
Definining the key points:
The Delaunay triangulation is performed to the averaged shape, and we get:
The mid-way face image is shown below:
As we can see the bottom half is blurred due to the mismatch in shape keypoints. As for the subset analysis, I generated two images of smiling faces and normal faces:
Again, we still see the blur effects. Then we first align the face (warping), and the resulting images become:
And for the subsets:
Before warping my face into the average geometry, I first generated the keypoint pairs as it follows:
Then, warping my face into the average geometry:
And when I warp the averaged face into my geometry:
This will be used as the base face shape for the extrapolation. The resulting gif is: