In addition to enabling robust matching, a key advantage of the invariant feature approach is that each match represents a hypothesis of the local 2D transformation. This fact enables efficient rejection of outliers using geometric constraints. We use broad-bin Hough transform clustering [1] to select matches that agree (within a large tolerance) upon a global similarity transform. Given a set of feature matches with relatively few outliers, we compute the fundamental matrix and use the epipolar constraint to reject remaining outliers.