CPS: 3D Compositional Part Segmentation through Grasping

Abstract

Most objects are composed of parts which have a semantic meaning. Ahandle can have many different shapes and can be present in quite different objects, but there is only one semantic meaning to a handle, which is "a part that is designed especially to be grasped by the hand". We introduce here a novel 3D algorithm named CPS for the decomposition of objects into their semantically meaningful parts. These meaningful parts are learned from experiments where robot grasps different objects. Objects are represented in compositional graph hierarchy where their parts are represented as the relationship between subparts, which are in turn represented based on the relationships between small adjacent regions. Unlike other compositional approaches, our method relies on learning semantically meaningful parts which are learned from grasping experience. This compositional part representation provides generalization for part segmentation. We evaluated our method in this respect, by training ion one dataset and evaluating it on another. We achieved on average78% part overlap accuracy for segmentation of novel part instances.