Extrapolate missing landmarks for expressive avatar

[kevinjosethomas]

Description

The expressive avatar loses one of its hands for many words in the database. This is primarily because Mediapipe failed to recognize the hands in the source video while augmenting the data, leading to the animation losing its hands for a few frames. This can probably be fixed by inferencing the location of the hands in missing frames by looking at its position in the last-visible and next-visible frames. I believe SignWave has an implementation of this here.