Scene Segmentation Based on NURBS Surface Fitting Metrics

This paper proposes a segmentation scheme jointly exploiting color and depth data within a recursive region splitting framework. A set of multi-dimensional vectors is built from color and depth data and the scene is segmented in two parts using normalized cuts spectral clustering. Then a NURBS model is fitted on each of the two parts and various metrics based on the surface fitting results are used to measure the plausibility that each segment represents a single surface or object. Segments that do not represent a single surface are recursively split in a tree-structured procedure until the final segmentation is obtained. Different metrics based on the fitting error and on the curvature of the fitted surfaces are presented and tested inside this framework. Experimental results show how a reliable scene segmentation can be obtained from this procedure.