The result of the matching procedure in a live stream from my laptop's webcam looks like this:

As you can see, most of the keypoints in the pattern image were matched correctly with their counterparts in the query image on the right. The printout of the pattern can now be slowly moved around, tilted, and turned. As long as all the corner points stay in the current frame, the homography matrix is updated accordingly and the outline of the pattern image is drawn correctly.

This works even if the printout is upside down, as shown here:

In all cases, the warped image brings the pattern image to an upright, centered position on the frontoparallel plane. This creates a cool effect of having the pattern image frozen in place in the center of the screen, while the surroundings twist and turn around it, like this:

In most cases, the warped image looks fairly accurate, as seen in the one earlier. If, for any reason, the algorithm accepts a wrong homography matrix that leads to an unreasonably warped image, then the algorithm will discard the outlier and recover within half a second (that is, within self.max_frames_no_success frames), leading to accurate and efficient tracking throughout.