XXL-CT Dataset Segmentation

The objective of XXL-CT dataset segmentation is to use machine learning to virtually divide 3D volumes of complete vehicles, acquired through XXL computer tomography, into their individual components. Gathering labeled training data for this type of data is challenging. Previously, entity classification from XXL-CT data required significant manual effort involving over 120 employees for several months. This chapter shows how to develop entity segmentation procedures which significantly reduce the time from measurement to virtual analysis. The most time-consuming part of the data processing chain is currently the segmentation of individual assemblies, especially when dealing with overlapping metal sheets. We aim to demonstrate the transferability of trained networks to different XXL-CT vehicle data, taking into account the large shape variations of different metal sheets and their contact, weld or rivet points. Challenges include low data quality which is affected by acquisition and reconstruction artifacts, a dataset size of up to 1.7 terabytes, and the large number of individual instances to be segmented and their interrelationships which must be taken into account. This contribution considers three major aspects: The segmentation of CT datasets by means of neural networks, the development of solutions for the annotation of XXL-CT data, and the transferability of the trained network.