Influence of Manual Inter-Observer Variability for the Performance of Deep Learning Models in Semantic Segmentation

Deep learning models for semantic segmentation have expanded their application span from general objects to structures of interest in real-world tasks. The current paper discusses one such practical challenge, i.e. the delineation of corrosion compounds on microscope images taken at the surface of unrestored archaeological artefacts. Apart from the complexity of learning irregular shapes, the influence of the human manual annotations on model training becomes pivotal. The aim of the current paper thus becomes to measure the impact of interobserver variability on the segmentation of four copper corrosion compounds. The results support the need for rigorous guidelines in the manual establishment of the ground truth.