The authors introduce a new and promising multivariate data clustering algorithm. They adopt a divisive strategy, that is, a procedure that partitions the input data space sequentially into a number of disjoint subregions.
After reviewing several well-known clustering techniques, namely the median-cut, mean-split, and k-means algorithms, they present their method. The clustering algorithm has to make two important decisions at each step while partitioning the input data space: first, which hyperbox should be partitioned and, second, which hyperplane is appropriate to subdivide the hyperbox. Both decisions are based on minimizing the sum-of-squared-errors.
Finally, the paper compares the performance of the algorithm and the above-mentioned clustering techniques on three collections of data of different dimensions from a color image database. The new clustering algorithm seems to perform better than the previously known methods.