Adaptive Feature Integration for Segmentation of 3D Data by Unsupervised Density Estimation

In this paper, a novel unsupervised approach for the segmentation of unorganized 3D points sets is proposed. The method derives by the mean shift clustering paradigm devoted to separate the modes of a multimodal density by using a kernel-based technique. Here, the attention is focused on the selection of the kernel bandwidth which typically strongly affects the level of accuracy of the segmentation results. In particular, a set of geometric features is computed from each 3D point of the given data. This set is projected onto a number of independent sub-spaces, each one associated to a different estimated feature, and overall forming a joint multidimensional (feature) space. In this space, we propose a method for selecting the best multidimensional kernel bandwidth in an automatic fashion, based on stability criteria. The final kernel considers each sub-space in an adaptive way in relation to the discrimination power of each feature, leading to accurate results when dealing with different types of 3D data