Normal Grouping Density Separation (NGDS): A Novel Object-Driven Indoor Point Cloud Partition Method

Precise segmentation/partition is an essential part of many point cloud processing strategies. In the state-of-the-art methods, either the number of clusters or expected supervoxel resolution needs to be carefully selected before segmentation. This makes these processes semi-supervised. The proposed Normal Grouping- Density Separation (NGDS) strategy, relying on both grouping normal vectors into cardinal directions and density-based separation, produces clusters of better (according to use quality measures) quality than current state-of-the-art methods for widely applied object-annotated indoor benchmark dataset. The method reaches, on average, lower under-segmentation error than VCCS (by 45.9pp), Lin et al. (by 14.8pp), and SSP (by 26.2pp). Another metric - achievable segmentation accuracy - yields 92.1% across the tested dataset what is higher value than VCCS (by 14pp), Lin et al. (by 3.8pp), and SSP (by 10.3pp). The experiment carried out indicates superiority of the proposed method as a partition/segmentation algorithm - a process being usually a preprocessing stage of many object detection workflows.