USING 3D MODELS TO GENERATE LABELS FOR PANOPTIC SEGMENTATION OF INDUSTRIAL SCENES

Abstract

Industrial companies often require complete inventories of their infrastructure. In many cases, a better inventory leads to a direct reduction of cost and uncertainty of engineering. While large scale panoramic surveys now allow these inventories to be performed remotely and reduce time on-site, the time and money required to visually segment the many types of components on thousands of high resolution panoramas can make the process infeasible. Recent studies have shown that deep learning techniques, namely deep neural networks, can accurately perform panoptic segmentation of things and stuff and hence be used to inventory the components of a picture. In order to train those deep architectures with specific industrial equipment, not available in public datasets, our approach uses an as-built 3D model of an industrial building to procedurally generate labels. Our results show that, despite the presence of errors during the generation of the dataset, our method is able to accurately perform panoptic segmentation on images of industrial scenes. In our testing, 80% of generated labels were correctly identified (non null intersection over union, i.e. true positive) by the panoptic segmentation, with great performance levels even for difficult classes, such as reflective heat insulators. We then visually investigated the 20% of true negative, and discovered that 80% were correctly segmented, but were counted as true negative because of errors in the dataset generation. Demonstrating this level of accuracy for panoptic segmentation on industrial panoramas for inventories also offers novel perspectives for 3D laser scan processing.