SASO: Joint 3D semantic‐instance segmentation via multi‐scale semantic association and salient point clustering optimization

Jointly performing semantic and instance segmentation of 3D point cloud remains a challenging task. In this work, novel framework called joint semantic-instance via multi-scale Semantic Association Salient clustering Optimization was proposed to tackle problem. Inspired by the inherent correlation among objects in space, Multi-scale (MSA) module explore constructive effect context information for is designed. For instance, segmentation, different from previous works utilising only inference procedure, Point Clustering (SPCO) put forward introduce algorithm into training phase, which impels network focus on points that are difficult be distinguished. Furthermore, affected structure indoor scenes, problem uneven distribution categories has rarely been considered but it significantly limits performance scene perception. To address issue, an adaptive Water Filling Sampling (WFS) balance category data presented. Extensive experiments variety changing datasets show authors’ method outperforms state-of-the-art methods both tasks segmentation.