Sun Chunyu ; Tong Xin ; Liu Yang

Abstract

Recognizing 3D part instances from a 3D point cloud is crucial for 3Dstructure and scene understanding. Several learning-based approaches usesemantic segmentation and instance center prediction as training tasks and failto further exploit the inherent relationship between shape semantics and partinstances. In this paper, we present a new method for 3D part instancesegmentation. Our method exploits semantic segmentation to fuse nonlocalinstance features, such as center prediction, and further enhances the fusionscheme in a multi- and cross-level way. We also propose a semantic regioncenter prediction task to train and leverage the prediction results to improvethe clustering of instance points. Our method outperforms existing methods witha large-margin improvement in the PartNet benchmark. We also demonstrate thatour feature fusion scheme can be applied to other existing methods to improvetheir performance in indoor scene instance segmentation tasks.