Investigating Feature Clustering for Generalised 3D Point Cloud Part Segmentation

We investigate the use of 3D part labelling networks for unseen object classes. Current methods for 3D part labelling require specific training for each object and choice of part labels. This makes the application of these networks to novel object classes difficult as labelled training data needs to be collected, and then the network (re-)trained for each new class. In this work we explore the clustering of features used to generate part labels in order to provide an (over-)segmentation of novel object classes. We propose a cluster-over-parts (CoP) metric to evaluate this approach. We find that between 75% and 95% of points can be successfully clustered by PointNet and PointNet++ on novel object classes. The most successful CSN network trained on a single class can accurately label 86.5% of points on average for unseen object classes, with individual unseen class accuracy from 76% to 95%.