Multi-Scale Point Cloud Analysis

Abstract

Surfaces sampled with point clouds often exhibit multi-scale properties due to the high variation between their feature size. Traditional shape analysis techniques usually rely on geometric descriptors able to characterize a point and its close neighborhood at multiple scale using smoothing kernels of varying radii. We propose to add a spatial regularization to these point-wise descriptors in two different ways. The first groups similar points in regions that are structured in a hierarchical graph. The graph is then simplified and processed to extract pertinent regions. The second performs a spatial gradient descent in order to highlight stable parts of the surface. We show two experiments focusing on planar and anisotropic feature areas respectively.