Medoid selection from sub-tree leaf nodes for k-medoid clustering-based hierarchical template tree construction

A method to construct a hierarchical template tree for pedestrian contour detection by iteratively applying a k-medoid clustering algorithm from the lowest level to the highest level was recently proposed and received much attention. Analysed here is the limitation of the method using lower level medoids as points of the next higher level, and proposed is a method of selecting a medoid from the leaf nodes of sub-trees corresponding to the lower level medoids.