The quality of the cross-correlation criterion depends on the amount of interesting signal present in the measurement area. Intuitively, if we try to cross-correlate two flat (blank) images, the returned measurement will have no meaning since any pixel could be a valid candidate.
The peak detector described below is used to localize bright features. Because the first frame in the stack has had less frames to estimate its sky background, it is usually more affected by noise and thus not so reliable to use for bright object detection. A good indicator of bright object position can be computed cheaply by taking the difference between the first 2 raw frames in input. This way, we have a frame with excellent sky and dark subtraction, but also negative objects. If we specify to the peak detector that it should not try to detect negative peaks, the same results are found as by using the first sky-subtracted frame, but since it is less noisy it is possible to find fainter objects, thus more correlating features.