Here are the results of using the Harris Interest Point Detector on my images.
In order to extract feature descriptors, I loop through the corners of an image and create 40x40 patches. I then resize it to 8x8 and normalize it. I finally return a list of all these descriptors. I do so for each image.
In order to match the feature descriptors, I calculate the distance between each pair of descriptors and if the fraction of the lowest two errors is smaller than 0.6 then I add this pair to my list of matchings.
After implementing the RANSAC algorithm, I got this homography matrix H: