Point-Level Label-Free Segmentation Framework for 3D Point Cloud Semantic Mining

Anan Du, Shuchao Pang, Mehmet A. Orgun

Abstract: 3D point cloud data semantic mining plays a key role in 3D scene understanding. Although recent point cloud semantic mining methods have achieved great success, they require large amounts of expensive manual annotated data. More importantly, the lack of large-scale annotated datasets limits those approaches in many real-world applications, especially for point-level semantic mining tasks such as point cloud semantic segmentation. In this work, we propose a novel point cloud segmentation framework, called Point-level Label-free Segmentation framework (PLS), that does not require point-level annotations. In this framework, the point cloud semantic mining task is formulated as a clustering problem based on mutual information. Meanwhile, our method can directly predict clusters that correspond to the given semantic classes in a single feed-forward pass of a neural network. We apply the proposed PLS to the shape part segmentation task. Experiments on the benchmark ShapeNetPart dataset demonstrate that our method has the ability to discover clusters that match semantic classes, and it can produce comparable results with methods using incomplete labels on several categories.