Investigation of Graph Neural Networks for Instance Segmentation of Industrial Point Cloud Data

Abstract

AbstractThe concept of Digital Twins (DTs) was introduced 15 years ago. There have been many research methodologies and software to implement DTs in different industries including manufacturing, construction, product design, and other fields. DTs help quantify operational risks, improve production time, assist predictive maintenance, enable real-time remote monitoring and thus better financial decision-making. The generation of DTs for existing industrial sites necessitates the use of laser scanners for the acquisition of point cloud data that capture the existing (as-is) conditions. Currently, human modelers manually segment point cloud data by overlaying 3D CAD models on top of the laser scans or validating laser scanned point clouds with 2D documentation and drawings. Our previous work has achieved effective point cloud processing with techniques such as instance segmentation and class segmentation of the collected and registered industrial point cloud data. Instance Segmentation is an important method of clearly partitioning each object to a human-understandable point cluster in complex laser scanned data, creating a Geometric Digital Twin of Industrial Facilities. The industrial point cloud data consists of pipes, valves, cylinders, and various other combinations of geometric shapes. Segmenting such data is a difficult task as the data is too complex to visualize and understand. In our previous work, CLOI-NET, which is the state-of-the-art architecture for instance segmentation of industrial point clouds, achieves instance segmentation with average accuracy of 70% using graph connectivity algorithms. This proved that there is scope for more accurate instance segmentation of complex industrial point cloud data with a focus on identifying topological connectivity between components of the point cloud (e.g., a piping network). Also, there have been many research methods, for instance segmentation in city-scale and indoor environments classifying cars, people, buildings, trees, roads, using different types of neural networks with satisfactory performance having average accuracy upto 85%. In this paper, we discuss the best algorithms/networks like Graphical Neural Networks and 3D-CNNs and how they can be used to perform instance segmentation of industrial data, which will eventually lead to a better version of DT implementation specifically for industrial point cloud data.KeywordsDigital TwinInstance segmentationGraphical Neural NetworkGeometric deep learningPoint cloudsComputer vision