Instance Segmentation of Industrial Point Cloud Data

The challenge that this paper addresses is how to efficiently minimize the cost and manual labor for automatically generating object oriented geometric digital twins (gDTs) of industrial facilities, so benefits provide even more value compared initial investment generate these models. Our previous work achieved current state-of-the-art class segmentation performance (75% average accuracy per point area under ROC curve, AUC, 90% in CLOI dataset classes) directly produces labelled clusters most important model objects (CLOI from laser scanned data. stands C-shapes, L-shapes, O-shapes, I-shapes their combinations. However, problem automated individual instances can then be used fit shapes remains unsolved. We argue use instance algorithms has theoretical potential output needed generation gDTs. solve through (1) using a CLOI-Instance graph connectivity algorithm segments an into instances, (2) boundary points improves Step 1. method was tested on benchmark segmented with 76.25% precision 70% recall among all classes. This proved it first segment cloud no prior knowledge other than label bedrock efficient gDT cluttered clouds.