Shape-Graph Based Object Recognition Using Node Context Embedding

Graphical object representation is fequently used for visual object recognition and detection methods.Since most machine learning methods requira vectorial input, significant research has been done on assigningfeature vectors to graphs - a process known as graph embedding. However, when one wishes to detect objects in alarger scene, it is a more viable strategy to assign feature vectors to graph nodes, and classify them individually. Inthis paper, we present a graph node embedding algorithm for 3D object detection based on primitive shape graphs.Our embedding algorithm encodes the local context of the selected node into the feature vector, thus improvingthe classification accuracy of nodes. The method also imposes no restriction on the structure of the graphs or theweights on the nodes and edges. The method presented here will be used as part of an intelligent object pairingalgorithm for Tangible Augmented Reality.