Fig. 2.

Stationary mean shift algorithm. In (a), the sample points are drawn on top of a kernel density estimate of the underlying distribution. At each iteration of the mean shift, samples from the neighborhood $S_h^{\prime }\left(X\right)$ are used to form the mean shift, $M_h^{\prime }\left(X\right)$. This is shown in (b) for t = 3 and a sample point $x_i^3$ marked by an “X”, with the points in $S_h^{\prime }\left(x_i^3\right)$ marked by open circles. The stationary algorithm converges more slowly than the original mean shift requiring 20 iterations shown in (d) as opposed to t = 8 in 1(d). There is no associated narrowing of the peaks as the stationary algorithm converges in contrast to 1(d).