It seems like they are registering the points-of-interest between frames/images so that they can create a consistently scaled set of output images even if you, for example, took a step further away from the filmed object. And then using these sort-of-normalized frames to interpolate between them.