Part-In-Whole Type 3d Partial Shape Retreival Based On Connected Faces With Pointnet Features

In this paper, we focus on “part-in-whole” type 3D partial shape retrieval, where we propose a new partial shape representation based on the idea of connected Faces accompanied with the features extracted from PointNet, given 3D shapes data as a solid model with Boundary Representation (BREP). The idea of connected Faces is inspired by the analogy with N-gram in Natural Language Processing. We conducted experiments with an ABC 3D dataset to compare several candidates of partial shape features, and to confirm that connected Faces approach has an advantage over traditional partial shape representations. Specifically, we compared the connected Faces approach of PointNet features with SNH (Surface Normal Histogram) as well as PFH (Point Feature Histogram), and found that connected Faces with PointNet outperformed other approaches in terms of several evaluation measures.