Associate Semantic-Instance Segmentation of 3D Point Clouds Based on Local Feature Extraction

Segmentation of 3D point cloud which can express the information of complex scene more accurately is an important basis for 3D scene understanding. However, how to effectively use 3D point cloud information for complex scenes is rarely discussed. This work proposes a two-stage network to achieve semantic segmentation and instance segmentation of point clouds. Specifically, a simple multitasking network is firstly developed by extracting the multi-category features of local point cloud, which can also achieve superior segmentation results. Then, a learnable network is established to make semantic segmentation and instance segmentation mutually promote each other, so as to segment complex scenes more accurately. The validity of this network is proved by experiments and evaluation on S3DIS dataset. Compared to other well-known networks, the proposed two-stage network shows its superiority.