Tracking-by-detection merging heuristic

[hbredin]

The heuristic used to merge trackers with detection (and trackers with trackers) is kind of messy and definitely needs some love.

Furthermore, not only is it messy but I already noticed some weird (and wrong!) behavior. See tracks 13 and 14 in the example available at http://nbviewer.jupyter.org/github/pyannote/pyannote-video/blob/master/doc/getting_started.ipynb that should definitely be merged into one track (but are not for some reason).

Before doing anything, the first thing to do would be to setup some kind of regression tests. In particular, this includes:

• manually annotating a few (multiple-people) videos (e.g. the one already used in the notebook documentation)
• implementing a tracking evaluation metric (ideally in pyannote.metrics but pyannote.video is fine)