Abstract

We point out a difference between the original mean-shift formulation of Fukunaga and Hostetler and the common variant in the computer vision community, namely whether the pairwise comparison is performed with the original or with the filtered image of the previous iteration. This leads to a new hybrid algorithm, called Color Mean Shift, that roughly speaking, treats color as Fukunaga’s algorithm and spa- tial coordinates as Comaniciu’s algorithm. We perform experiments to evaluate how different kernel functions and color spaces affect the final filtering and segmentation results, and the computational speed, using the Berkeley and Weizmann segmentation databases. We conclude that the new method gives better results than existing mean shift ones on four standard comparison measures ((cid:2) 15%, 22% improvement on RAND and BDE measures respectively for color images), with slightly higher run- ning times ((cid:2) 10%). Overall, the new method produces segmentations comparable in quality to the ones obtained with current state of the art segmentation algorithms.