Shape Similarity Based on a Treelet Kernel with Edition

Several shape similarity measures, based on shape skeletons, are designed in the context of graph kernels. State-of-the-art kernels act on bags of walks, paths or trails which decompose the skeleton graph, and take into account structural noise through edition mechanisms. However, these approaches fail to capture the complexity of junctions inside skeleton graphs due to the linearity of the patterns. To overcome this drawback, tree patterns embedded in the plane have been proposed to decompose the skeleton graphs. In this paper, we reinforce the behaviour of kernel based on tree patterns by explictly incorporating an edition mechanism adapted to tree patterns.