This seems like a simpler version of what Microsoft demonstrated in the paper. Between the final and original shots in Instigram's demos you can still see much of the original motion. This is especially apparent in the car clip.

What it looks like the Instagram app is doing is selecting and stabilizing individual frames instead of the full 3D scene reconstruction that the MSFT paper described. EG there's no pop in and out of details like the MSFT demos and there's still a small but noticeable judder in some of the videos.