I'm a bit confused about the intuition behind this algorithm, as well as the details of the algorithm. I am mostly confused about why both the endpoints and the length of the field are necessary. It seems to me that only two of these are necessary, and that the third should be very dependent on the other two. This algorithm would make more sense to me if, instead of finding the fore detectors independently of the aft detectors and then using the distribution of the lengths of fields to pair them together, maybe just the fore detectors and the lengths were calculated, or just the fore and aft detectors were calculated together.

Perhaps my confusion with the intuition behind this algorithm stems from my confusion with the actual LearnDetector Weak Learner. As I understand it, the LearnnDetector goes through all possible boundaries. For each boundary, it determines the "best" length before the boundary to match a field and the "best" length after the boundary to match a field, for that boundary (I am unclear about this as BestPreExt and BestSufExt are not described). It then compares this to the previous best matches, and saves the current one if it is "better" than the previous best matches. If this is the case, it seems like the algorithm is finding the start and end of a field at the same time, instead of just unconnected fore and aft boundaries. Why doesn't the LearnDetector algorithm just return the best of p and s that it finds?