October 5th 2015, 10:30am in room Byron blanc
Abstract: Over the last decades, the evolution of acquisition techniques yields the generalization of detailed 3D objects, represented as large point clouds composed of millions of vertices. The complexity of the involved data often requires to analyze them for the extraction and characterization of pertinent structures, which are potentially defined at multiple scales.
During this talk I will present a point-based multi-scale analysis framework, inspired by the Scale-Space theory. Its core component is the Growing Least Squares (GLS) descriptor, which measures the derivatives of a fitted primitive at increasing scales. I will present how this descriptor can be used to decompose a point cloud as a set of geometric structures defined at multiple scales using unsupervised clustering techniques. Recent results also shows that the GLS descriptor can be used to estimate a relative scale factor between two points, for instance to register point clouds acquired at arbitrary scales (e.g. photogrammetry output). In addition to scale analysis, I will present on-going research based on the analysis of primitives in space, and perspectives on point-based Scale-Space representations.
