A computational strategy used in optical flow estimation. By the way, as far as I can see, it could be applied to any other problem with similar assumptions.

The basic idea is: optical flow assumes that the displacement vectors between two images are small enough to use effectively a linear approximation (truncating a Taylor expansion) of the optical flow equation. This is not always true, obviously.

Then, a good solution is to try a **coarse-to-fine** approach, that is (simplified in two step):

- solve the problem at a lower resolution, where the difference between the two images is less evident;
- use the coarse result to reduce the difference at the higher resolution;
- solve the problem at the higher resolution for fine-tuning;

Practically speaking, all this is achieved building a multi-resolution pyramid (gaussian or laplacian) of both the two images, and starting to apply the optical flow from the bottom, then climbing the pyramid. At each step, the next image is pre-warped with the last o.f., until you get the final result.