Annotator based on optical flow about dvt

[statsmaths]

Adding optical flow would be a great contribution if you're willing to put it together!
It is possible to compute the optical flow for the last frame because each batch
includes a look-ahead region (of equal size). That, for example, is how we are able
to do difference annotations. It is also possible to carry-over the key points to the
next batch using the cache attribute of the annotator. It doesn't seem like there would
be much wasted computation because even at a very high cut rate, cuts should only
be happen <5% of the time.

You are entirely correct that sometimes it is best to do a first past to get, e.g., cuts and
then run a second pass after having detected those cuts. This is entirely possible within
DVT. For example, you could

• run one pass with a cut annotator
• aggregate the cuts to get the shot breaks
• feed the shot breaks to initiate a new frame annotator
• re-run the pipeline

What's nice about that structure is that it is even possible to use a different program to
detect or manually specify the shot breaks before doing the optical flow.

[Nanne]

I didn't notice the look-ahead region, that's a clever solution! I'll have a go at adding optical flow then, as the structure definitely sounds flexible enough to make it work in a nice way, thanks!