{Raymond Ptucha Shagan Sah Saloni Jain Atir Petkar Rohan Dhamdhere Miguel Dominguez}

Abstract

Depth sensors used in autonomous driving and gamingsystems often report back 3D point clouds. The lack ofstructure from these sensors does not allow these systemsto take advantage of recent advances in convolutional neural networks which are dependent upon traditional filteringand pooling operations. Analogous to image based convolutional architectures, recently introduced graph based architectures afford similar filtering and pooling operationson arbitrary graphs. We adopt these graph based methodsto 3D point clouds to introduce a generic vector representation of 3D graphs, we call graph 3D (G3D). We believewe are the first to use large scale transfer learning on 3Dpoint cloud data and demonstrate the discriminant powerof our salient latent representation of 3D point clouds onunforeseen test sets. By using our G3D network (G3DNet)as a feature extractor, and then pairing G3D feature vectorswith a standard classifier, we achieve the best accuracy onModelNet10 (93.1%) and ModelNet 40 (91.7%) for a graphnetwork, and comparable performance on the Sydney Urban Objects dataset to other methods. This general-purposefeature extractor can be used as an off-the-shelf componentin other 3D scene understanding or object tracking works.