If you watch the George Bush segment, it shows the reconstructed video from a different angle (straight-on, eliminating the head movement).

> We note that our method does not guarantee an accurate fit to ground-truth geometry, as the shape of the face may change in each frame and single-image cues are not sufficient for this purpose. Rather, we seek to produce a reasonably convincing model (leveraging all available imagery) which optimally fits the image information in each frame.

This makes it seem like the mesh produced for each frame is specially deformed to render optimally for that camera angle, and it wouldn't necessarily look perfect from other angles.