Hierarchically Distributed Dynamic Mean Shift

A fast and memory-efficient method is presented for dynamic mean shift (DMS) algorithm, which is an iterative mode-seeking algorithm. The DMS algorithm requires a large amount of memory to run because it dynamically updates all samples during the iterations. Therefore, it is difficult to use the DMS for clustering a large set of samples. The difficulty of the DMS is solved by partitioning a set of samples into subsets hierarchically, and the resultant procedure is called the hierarchically distributed DMS (HDDMS). Experimental results on image segmentation show that the HDDMS requires less memory than that of the DMS.