Thanks for the feedback ! I'm planning to use more robust and scale invariant features for this so this problem of different levels of zoom would be solved (at least to the limits of what can be done).