Classical approaches of feature line detection rely on curvature derivatives. They generally suffer from a common problem: the connectivity is hard to obtain and it is impossible to generate intersections between feature lines. This article presents a method to extract feature lines on 3D meshes. In order to sort out the recurrent issues of traditional approaches, we propose a novel algorithm based on two ideas. First, all the mesh vertices are marked according to the curvature values: a binary map with candidate regions is then constructed. The second idea is to isolate each candidate region and transform it into a line. To achieve this, we parameterize the region into its 2D regular representation. We then perform a skeletonization to obtain lines with high connectivity. By applying the inverse parameterization, the feature lines are mapped back onto the 3D mesh. In the end, we extract perceptual salient parts and above all connected feature lines. In order to evaluate and validate our algorithm, we compare our method to classical ones and apply our technique to a geological context.
Feature line detection of noisy triangulated CSGbased objects using deep learning
