The project was to take separate red, green, and blue image channels,
created by taking black and white images, and combine them to yield full
color images. It was assumed that the channels were only translated relative to each other.
The approach I took was to pyramid the image down to about 200x200, then do an search over the window [-10, +10]
in both x and y directions. From that estimate, each successive level of the pyramid was used to refine the estimate,
over the window [-1, +1]. An issue with this approach was errors at the edge of the image caused the matching to fail.
Using a smaller window in the center of the image fixed this issue. A second issue was a few matches failed to match correctly
because of lighting, when using a sum square differences matcher. This was fixed by using a normalized cross correlation matcher.
The match on image 87u failed, likely due to the highly repetitive pattern of the hut that fills most of the image.
The mismatch is shifted by the size of the repeating pattern, indicating that the matcher failed at one level of the pyramid,
and then propagated the error, finding the next best match.