Did they cherry-pick the examples? It looks suspiciously good.
That's what I was wondering. This is way beyond the level of accuracy that I have seen from anything else.
Honestly, I find this practice unacceptable, and unfortunately I see it a lot. The academic world should frown upon this kind of presentation.
It isn't clear from the paper but their appearance model shown in the video segmentation demo is likely the same thing. It shows poor performance at times (see the train at 0:32).