Probabilistic Boundary-Guided Point Cloud Primitive Segmentation Network.

Three-dimensional point cloud primitive segmentation aims to segment an original entire point cloud into a set of geometric primitives with different types, which is widely used in the manufacturing industry. The primitive segmentation task is challenging faced with complex shapes and ambiguous boundaries. We observe that the boundary properties of point clouds have not been fully investigated and exploited in previous works, and the primitive segmentation performance is not satisfactory near boundaries, especially gradually-changed boundaries. In this article, we propose a novel probabilistic boundary-guided primitive segmentation (PBPS) network to improve the primitive segmentation ability by emphasizing the boundary cues. First, the point cloud boundary is represented by Gaussian distribution instead of binary representation, which can describe boundaries more informatively and also provides an indication of the ambiguous relationship between point cloud boundaries and inner regions. Second, a probabilistic boundary-guided feature fusion (PBFF) module as well as an instance clustering and type voting strategy are proposed, which process the boundary points and nonboundary points conditioning on different boundary probabilities, to reduce the impact of boundary ambiguity on primitive segmentation. Third, a primitive instance contrastive loss is designed which can relatively loosen the constraints on the distances from boundary points to the centroid in the embedding space. The effectiveness of PBPS was verified by a series of experiments on two computer-aided design (CAD)-based datasets and two real-scene datasets.