Structure-aware shape correspondence network for 3D shape synthesis

Abstract Structure-aware methods can be used to create new 3D shapes by reusing existing parts of given shapes. However, it is still a challenging task to obtain the corresponding parts among different shapes using structure information. We propose a structure-aware shape correspondence network, SASCNet, which can be used to obtain the corresponding parts between a pair of over-segmented shapes and then perform 3D shape synthesis effectively. In this network, the structure features of parts, which are extracted from the structural graph of a shape by the graph attentional layer, can be further used to calculate the part correspondence matrix and obtain the corresponding parts by the correspondence module. According to our part reshuffle experiments on several pairs of shapes, reasonable new shapes are created effectively. Furthermore, the part correspondence performance of our SASCNet is verified by comparing several correspondence results with that of two reported methods. Our approach is found to achieve better performance than theirs in some experiments.