CVPR Daily - Monday

So much data these days comes from scans, and that scanned information from the real world appears as point clouds . This work takes an input point cloud and finds a triangulation between the points to give a watertight and manifold surface . In computational geometry there are great classical tools that guarantee well- shaped triangles and watertight triangulation to create a high-quality mesh . However, these classical methods are not based on learning, so strict conditions such as finetuning of parameters are needed to make them work. In this paper, Marie-Julie mixes the best of both worlds, combining existing methods with learning methods that offer more flexibility . Many learning-based methods work on point clouds, like PointNet, or generative models. They can adapt to different kinds of meshes and different sampling conditions to build the triangulation from the point cloud. “ Point clouds are flawed in the sense that you don’t know what the relation is between the different points, ” Marie-Julie explains. “ If you have a surface and three points, you don’t know if they will be connected between each other. When you use classical methods, they don’t know this information. With our method, you can easily augment an existing method for generating point clouds so that it will learn on the surfaces. ” The work uses a patch-based method . Given a patch of point cloud, the aim is to find the corresponding 2D parameterizations. It takes a patch and flattens it in 2D using a network inspired by FoldingNet . Once they have this 2D parameterization, the team compute the Delaunay surface element , which is based on computing the Delaunay triangulation from the 2D parameterization. This gives a surface element that is manifold and watertight locally by construction. Marie-Julie Rakotosaona is a third-year PhD student at École Polytechnique in Palaiseau, France under the supervision of Maks Ovsjanikov. Like Vicky Kalogeiton, Marie-Julie is part of GeoViC, a Computer Graphics group tackling geometric modeling and computer animation problems. Her paper on reconstructing surfaces from a point cloud is one of the first to be presented at this year’s conference. She speaks to us ahead of its debut today. Learning Delaunay Surface Elements for Mesh Reconstruction 14 DAILY CVPR Monday Presentation