We choose the top n as the positive examples in the
query set and rerun the MR algorithm to refine the
retrieval result. The vector Y will have more nonzero
components that will spread their ranking scores. Thus,
the sequence ff .t /g converges to
where yj = 1 if xj is a typical frame in the positive
example set (otherwise, it is 0) and r is the number of
positive examples. In this way, an RF procedure without