Adaptive over-relaxed mean shift
Chunhua Shen, Michael J. Brooks · 2006
Mean shift is a popular nonparametric density estimation method. One of its drawbacks is that it converges slowly in many cases. Inspired by the successful accelerated variants of bound optimisation algorithms such as Expectation Maximisation, we propose an accelerated version of the mean shift algorithm. The appealing property is that, compared with the standard mean shift algorithm, the number of iterations to convergence is signicantly decreased. Additionally for the Gaussian kernel, no extra computation is introduced at each iteration. We empirically show on various data sets that the extended algorithm can provide a considerable speedup.