I have an idea of a new way to compress movies. Imagine a short movie of man who smiles. It is something like 40 frames long (in the picture below you see only first, middle and last frame, just for simplicity).
Now imagine a person marks some control points on the first frame...
... and then marks location of that control points on all frames.
Now the encoder bends all the frames, so control points come back to their starting position.
Now it encodes the movie, using some normal method (eg. xvid). It also saves coordinates of control points on all frames.
When it's time do decode the movie, decoder decodes it...
And then – using the information about control points coordinates – re-bends frames.
We can watch our movie.
I believe this method (which is a special sort of move prediction) will give good results. Because of frames bending the differences between consecutive frames will be less, so the movie will encode better, and remembering coordinates of control points will not take so much of precious bytes. Maybe it would be also possible that it will be computer, not a human, who marks control points on frames.