Image Analysis, Classification and Change Detection in Remote Sensing, 3rd Edition

Exercises 363

A direct implementation of this algorithm would be prohibitively slow, since

mean shifts a re to be computed for all of the pixels. For a partial solution in

Python using the OpenCV package, see Exercise 12(c).

In the ENVI/IDL extension MEAN_SHIFT_RUN described in Appendix C, two

approximations are employed to sp eed things up. First, all pixels within ra dius

r of convergence point w are assumed also to converge to that mode a nd are

excluded from further processing. Second, all pixels lying sufficiently close

(within a distance r/2) to the path tr aversed from the initial vector y

ν

to the

mode w are similarly assigned to that mode and