0n convergence of the mean shift algorithm
Tzon-Liang Shieh, Jiarui Zhang, Shih-Yu Chiu, Leu-Shing Lan · 2008
As a nonparametric statistical method, the mean shift algorithm has recently attracted much attention in the computer vision community due to its efficiency in motion tracking and clustering analysis. Although convergence of the mean shift algorithm has already been proved, there are still some pitfalls in its convergence behavior which remain unobserved. In this work we investigate the premature convergence phenomenon of the mentioned algorithm. Two necessary conditions to examine premature convergence are analytically derived. We give some examples to confirm the correctness of the proposed theorems.