Exploiting Local Geometry for Feature and Graph Construction for Better 3D Point Cloud Processing with Graph Neural Networks

We propose simple yet effective improvements in point representations and\nlocal neighborhood graph construction within the general framework of graph\nneural networks (GNNs) for 3D point cloud processing. As a first contribution,\nwe propose to augment the vertex representations with important local geometric\ninformation of the points, followed by nonlinear projection using a MLP. As a\nsecond contribution, we propose to improve the graph construction for GNNs for\n3D point clouds. The existing methods work with a k-nn based approach for\nconstructing the local neighborhood graph. We argue that it might lead to\nreduction in coverage in case of dense sampling by sensors in some regions of\nthe scene. The proposed methods aims to counter such problems and improve\ncoverage in such cases. As the traditional GNNs were designed to work with\ngeneral graphs, where vertices may have no geometric interpretations, we see\nboth our proposals as augmenting the general graphs to incorporate the\ngeometric nature of 3D point clouds. While being simple, we demonstrate with\nmultiple challenging benchmarks, with relatively clean CAD models, as well as\nwith real world noisy scans, that the proposed method achieves state of the art\nresults on benchmarks for 3D classification (ModelNet40) , part segmentation\n(ShapeNet) and semantic segmentation (Stanford 3D Indoor Scenes Dataset). We\nalso show that the proposed network achieves faster training convergence, i.e.\n~40% less epochs for classification. The project details are available at\nhttps://siddharthsrivastava.github.io/publication/geomgcnn/\n