Examine and tune the depth-proportioned cropping algorithm

[jonfroehlich]

Per our discussion here (#2 (comment)), we may want to re-examine our auto-cropping approach. Would require a ground-truth bounding box dataset (could still use Tohme as @tongning did?).

[galenweld]

In particular, we're discussing tuning the crop-size-based-on-depth aspect of our auto-cropping algorithm. This should improve quality of the crops.

@kaviMD has already made lots of improvements wrt increasing the speed of the crops. I'm renaming this issue to make it more clear.