Visualization and handling of real annotation variation in AI-supported image analysis

• Applicant: Professor Dr.-Ing. Volker Rodehorst
• Subject Area: Structural Engineering, Building Informatics and Construction Operation
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 556314568

As digitalization progresses, it is of great importance that we can trust the systems used in practice, especially when inspecting critical infrastructure such as bridges. Supervised machine learning processes depend crucially on the quality of the data used. Large models require extensive and high-quality data in order to be trained effectively. If the data is insufficient or of poor quality, the models learn incorrect information, which significantly limits their practical applicability. Annotation variations (AV) are an important problem in damage detection. They include uncertainties and errors that arise during the annotation process and are adopted by the neural networks. This leads to an overfitting of the models to these errors, which reduces their performance and reliability. AV can occur in various forms: different interpretations of damage delineation, inconsistencies between annotators in classification, overlooking of damage, marking of areas without actual damage and other misinterpretations. These variations result from the complexity of the damage patterns and the challenge of defining clear annotation guidelines. The goal of VarInSightAI is to investigate and address the impact of AV on damage detection. This includes identifying, quantifying, analyzing the impact and correcting AV. In addition, a data set with minimal AV will be created and methods for dealing with AV during training and evaluation will be developed. These approaches aim to improve the reliability and accuracy of machine learning models, especially in critical applications such as structural monitoring.

• DFG Programme: Research Grants